HN Simulatornew | past | comments | lists | submit | aseipp's commentslogin

It's extremely funny how many of your replies are proving your point exactly. One reply says "Yours is easy to hold since it can't have evidence" when it is simply stating that _there is no evidence of a claim_. Imagine that, "There's no proof of X" is somehow a magically unfair argument because it can never be proven! Wow! Nobody has ever thought of this before. Guess we have to assume there's inherently merit to X, then, and don't have to worry about actually constructing a proof of it (that requires patience and knowledge and first-party sources and does not get you internet points).

The entire article that thread references never mentions any hardware, at all, only their software platform. The "Watchdog" referred to in the (editorialized?) title is literally never mentioned, and it is not really a piece of "hardware", it is a piece of _software_ called "Sentry" that (clarified in the Nvidia blog) they run on Bluefield-4 hardware (the "chip") which is their network card. It only exists in $10mil+ clusters. Just an awful and misleading article and headline, and an awful thread filled with people who know nothing about how hardware works arguing over nothing.

Most of the hardware functionality from a GPU that models need is generic and used every time you use a GPU for any compute. You can implement language models in a GLSL shader; modern GPUs have been programmable for 20 years, that's the point of their existence. The best thing they could do is just literally nerf the consumer hardware substantially to make tokens unusable or stop selling consumer hardware. They don't need spyware to do that. It would also make their products dogshit.

More broadly, "Compute" is basically a resource that our economy relies on to a large degree, it has been that way for decades. And it turns out it's very bad when you just destroy the thing your economy relies on. And doing so doesn't solve problems like "Current computers exist today" or "Other people make computers". So who gives a shit about Jensen? Comments here talking about "taking him out"? What's actually more likely is just that really powerful people today are often corrupt and also really fucking stupid. He played a stupid game and lost.

None of this stuff is putting the current open weights genie back in the bottle IMO. Amodei using RoboStormtrooperGruppen9000 bots and going door to door to take peoples GPUs at gunpoint is a more plausible method of doing so.


Yeah, even setting aside the impact and implication on the AI-tech-driven stock market of such a mandate, I don't think it's technically doable to implement either. What if you modified/wrote your own inference engine which uses a somehow slightly modified version of model weights, how would you want to detect (at a firmware level) that compute to be an "unsigned" model over, I don't know, a video game?

Open-weight models, at an absolute minimum the ones that already exist, are here to stay, and will continue ([0]) to get drastically cheaper to run.

[0] I shall continue to reference one my favorite articles in recent months: https://jyn.dev/tokens-too-cheap-to-meter/


"Things change over time" that is how the linear passage of time tends to work, yes.

The OS absolutely controls functionality of the baseband, it would be a legal nightmare if it just didn't actually respect proper radio settings. But from a design perspective, it doesn't actually matter that much at all I feel. Any practical and secure design for a cellular phone has to assume the network is totally compromised and hostile, and so it can never be in the trusted base. From this perspective, adding the radio chip and its software to that category is not actually a huge leap. Even today vendors who own all the IP on a bus will still insert IOMMUs, etc inbetween components. It isn't dissimilar to game consoles; they assume that anything outside their die is compromised and hostile, and if you start from that premise, you can still make it very secure even when your threat model includes hostile components on the PCB.

Radio firmware is probably somewhat broken in practice (what isn't these days) but ultimately the openness stuff is a red herring IMO as far as security is concerned, you either assume it's compromised or not.


You can buy Agilex devices for like $100 now. They're Altera's primary chip family at this point in every device range. You can also use Quartus for free on these devices and even download it all through Docker. Knock yourself out.

The reality is just that it's a lot of work, and Agilex is probably an order of magnitude (or two orders) more complex in scope than any existing FPGA with reversed or open specs. It's going to require a lot of expertise even with superhuman assistance. Even less "quirky" but fully open source designs like GateMate have very poor fmax and utilization in my experience. If people are going to Claude shit out in the EDA space I'd frankly rather have them work on more important day to day mundane things that are still lacking but important, eg improved performance, power analysis/floorplanning, sdc/constraint based clock synthesis and related tools, etc.

And yes, breaking your device is also a risk for some features too (eg Lattice ECP5 bitstream encryption has never been RE'd since it requires blowing fuses.) But the actual problem requires far more expertise than having Claude reverse engineer gameboy games.


> Even less "quirky" but fully open source designs like GateMate have very poor fmax and utilization in my experience.

The terrible Fmax on GateMate is specifically a result of its "quirky" design. There's no dedicated clock network on the part - clocks have to be routed through fabric just like everything else. Running those signals everywhere consumes routing capacity and incurs delays; the resulting clock skew is awful.


That type of design seems highly resistant to classic fmax raising techniques like pipelining because pipelining can't fix clock skew.

A pipeline can tolerate clock offsets. E.g., an ALU like the Pentium 4's staggered ALU but within "a single clock cycle" might allow the more significant input bits to arrive at a time offset relative to the time for the first bits of inputs. Parsing variable length encoded instructions has a similar "carry propagation". Similarly, different stages might use different amounts of the cycle. Grid network routing can also localize time synchronization.

Getting the clock routing delay to approximate the delay that is designed to be tolerated seems likely to be rather tricky, but I think clock offsets from imbalanced travel distance do not have to make all of that time wasted.

I might well be wrong as I am only aware of some very general concepts of hardware design. Even in my ignorance, I am aware that theoretically possible in some cases is not the same as practical/worthwhile in general.


Subverting the secure boot/upgrade chain would require a SecureROM exploit and presumably a compromise of the Secure Enclave, so the answer is almost certainly "no" (even prior major exploits from recent memory like usbliter8 are tethered and don't exploit the enclave; the last true untethered Jailbreak was iOS 9 or something.) An upgrade would fix it. The internet reporting claims this break they are showing in this video is untethered though, which is extremely rare on its own.


Embedded Swift is honestly very cool. I never really used Swift itself before, but I think it's a good niche since embedded dev could stand for another tool and the language is pretty nice. It also doesn't really suffer from most of the issues people have historically brought up with Swift (cross platform, slow compiles) etc as it is so small, at least to me. :)

If you're starting from scratch I've found you can actually write ~everything in Swift quite easily (reset vectors, compiler emitted intrinsics, etc). But beyond that I think the best feature is that the C/C++ interop works really well and comes out of the box. Makes two-way adoption easy. I wrote a small boot firmware for a simulator and for fun decided to add signatures to the boot process (mldsa44+jq255). You can just copy the .c and .h files and import them into a modulemap file and you're off to the races. You can just get simple .o files and pass them to the linker. There are only 4-5 "freestanding" functions you have to implement. You can turn off heap allocation. Etc.

It would probably be pretty cool and quite easy for instance to write Zephyr or U-Boot drivers, etc in Swift using these techniques.


Can't you just keep using Mold2 for the initial toolchain bootstrap and then build mold3 with Rust as part of the "final" set of toolchains that are to be used to compile everything beyond that? It isn't perfect but it's congruent with how many other components in most open reproducible bootstrapping flows work; ie using GCC 4.4 or whatever just to bootstrap GCC 10. It's annoying but it only needs to be done once and then you just use it to bootstrap newer components forever. Doesn't help with compile times, though.


I would agree if not for the fact rust depends on almost the entire tree, so doing what you propose would increase our full tree verification time from 6ish hours to 12ish hours erasing any wins mold was giving us in the first place.


I may be misinterpreting the parents point, but you're not building E.g. tinycc for it to be a final artefact; why not continue to build mold2 in its current place in the build chain? It's not going anywhere, much like tinycc isn't.


We have to rebuild tinycc any time a dependency of tinycc changes. Likewise we would have to rebuild mold every time a dependency of mold changes, including rust and all dependencies of rust. It is a problem of almost full tree dependency recursion.


Yes. I'm saying that mold3 gets added to the end of the build tree, and mold2 remains where it is currently, which doesn't change how often you need to build those and let's you continue to use mold2 to speed up the builds (which I've now seen is what you'll likely do).

My point though was that you already build many end-of-life projects just so they can be used in the build, mold2 can now one of those rather than being built for its use in the final distro.


I indicated as much in my original comments. Pinning mold2 and backporting relevant fixes to it forever is our responsibility now that upstream has abandoned the linux-distro-bootstrapping use case. It is just unfortunate as that was supposed to be one of the goals.


Frozen code seems even better for this use case.


why can't you do:

C/C++ + libc ↓ LLVM/LLD ↓ mrustc + minicargo ↓ rustc 1.90 Cargo 1.90 ──→ OpenSSL ↓ Python x.py ↓ Rust 1.91+


That is what we do, and it takes several hours on most machines. Having to do that once to get mold and a second time to then build with the benefits of mold defeats any advantages for us.


You could also make the alternative argument: the internet as a truly decentralized network as imagined by nerds was never going to actually work due to its (now obvious) impact on the political economy of the world and its actors, and because fundamentally centralized economies of scale are how humans tend to organize society. So why are internet nerds on forums so narrow minded that they don't read understand this, because they don't read books? But both of these "arguments" are pointless posts designed to get head-pats, because everyone has made up their mind. Buy your DV certs from another ACME provider.


Someone from Qualcomm posted a "DT-ACPI Hybrid Mode" set of patches a while back that let you call ACPI table functions, etc even while booting with the device tree, allowing you to control some things like keyboard lightning/power management independently. That would be a good first step but I'm not sure it went anywhere in the meantime. In theory supporting Device Tree by including the blob in a ROM somewhere shouldn't be much more work (if any) than including working ACPI tables, but we all know how that works out in practice!


> supporting Device Tree by including the blob in a ROM somewhere shouldn't be much more work (if any) than including working ACPI tables, but we all know how that works out in practice!

This is definitely a thing in some embedded systems I’ve used. There are some Marvell boards which do this. They all use U-boot though.


The problem is that the content of the device tree isn't really stable -- if a new kernel version has a driver with a new required parameter for instance, suddenly your ROM device tree is out of date and much less useful.


Oh, that's a good point I forgot about; I suppose DT interfaces aren't considered part of the stability contract for users... Really Device Tree is somewhat "Linux-specific"-enough and low level enough that issues like this aren't ideal. But if we could just get to the point that was a problem, I'd be happier than now, I suspect!


In theory device trees are supposed to be a OS-neutral stable ABI. Being able to put a device tree into unchangeable mask ROM is an often mentioned design goal.

However, there is a fatal flaw in the process that makes this a pipe dream in reality. In Linux, these rules only apply to reviewed and merged drivers. And during review any new DT interface will almost certainly be bikeshedded in some way.

So unless your chip vendor has upstreamed ALL their drivers before a device is manufactured, the device's firmware-provided DT will very likely work only against a vendor kernel and requires updating.


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

Search: