atq2119

Datacenter GPU dies cannot be binned for Geforce because they lack fixed function graphics features. Raytracing acceleration in particular must be non-trivial area that you wouldn't want to spend on a datacenter die. Not to mention the data fabric is probably pretty different.

> once you open the door, the situation resets

That's the root cause error of your thinking.

The prizes aren't reshuffled and the host's choice of doors depends on both the player's choice and on information that is hidden to the player. No way you can treat that as a reset.

> HPC platforms

GPUs don't reorder instructions at all.

You don't really want or need classical unit tests for compilers. You need a harness that you can feed a program representation into and get a transformed one out, then check the result -- ideally mostly just by running it, or alternatively by comparing the output against a reference.

A lot of this is personal preference, but I still feel like the most memorable shows tend to be the ones that have a bit of both. Season-long stories, but also episodes that can stand on their own.

In a show like Stranger Things, almost none of the episodes are individually memorable or watchable on their own. They depend too much on the surrounding episodes.

Compare to e.g. Strange New Worlds, which tells large stories over the course of a season, but each episode is also a self-contained story. Which in turn allows for more variety and an overall richer experience, since you can have individual episodes experiment with wacky deviations from the norm of the show. Not all of those experiments will land for everybody (musical episodes tend to be quite divisive, for example), but there is a density to the experience that a lot of modern TV lacks.

As usual, the problem is: how fast does this happen?

I've seen some weird breakage recently in vscode. The C++ support failing to parse sources correctly (for LLVM), Rust debugging no longer showing vectors properly. Not sure if this is some bizarre interaction with my setup (which is pretty vanilla Ubuntu) or a regression in basic functionality brought on by an over-emphasis on AI features.

It is worrying that for many months now, pretty much all the content of changelogs has been about AI.

The benefit is also RAM bandwidth. That probably adds to the confusion, but it matters a lot for decode. But yes, RAM capacity requirements stay the same.

This framing is extremely counterproductive, though.

Most societal problems cannot be fixed entirely. There will always be child sex abuse just like there will always be murder, theft, tax evasion, and drunk driving. It makes sense to see if things can be improved, but any action proposed must be weighed against its downsides. Continued action by police is a good thing, but laws for that have been established for a long time, and the correct answer may well be that no further change to laws is required or appropriate.

(Ab)using child sex abuse to push through surveillance overreach is particularly egregious considering that by all objective accounts most of it seems to happen in the real world among friends and family, without any connection to the internet.

Those languages don't treat blocks as expressions, so you really can't do the same thing there. Something very similar, yes. But not the same.

Not mentioned in the article but kinda neat: you can label such a block and break out of it, too! The break takes an argument that becomes the value of the block that is broken out of.

It's a silly offhand remark at the end of the article, but anybody who is genuinely interested in whether they've been tying their shoes wrong will enjoy Ian's shoelace site: https://www.fieggen.com/shoelace/

Eh, I get it. I also independently came up with these MPMC approaches before hearing about LMAX. They're not entirely trivial but they do follow fairly naturally from the problem space when you really think about it. It's a good piece of engineering, but the one thing that's really unique about LMAX is the amount of advertising it gets.

Looks like a good write up, but I'd caution that some of the statements about memory models aren't completely accurate.

The terms relaxed, acquire, and release refer to how an atomic operation is ordered against other accesses to memory.

Counter to what the article states, a relaxed atomic is still atomic, meaning that it cannot tear and, for RMW atomic, no other access can go between the read and the write. But a relaxed atomic does not order other accesses, which can lead to unintuitive outcomes.

By contrast, once you've observed another thread's release store with an acquire load, you're guaranteed that your subsequent memory accesses "happen after" all of the other thread's accesses from before that release store -- which is what you'd intuitively expect, it's just that in modern systems (which are really highly distributed systems even on a single chip) there's a cost to establishing this kind of guarantee, which is why you can opt out of it with relaxed atomics if you know what you're doing.

I'm always skeptical of claims like these. Ultimately, programming is about writing down instructions for computers to follow. Existing programming languages evolved to be ergonomic for that, proof assistants less so.

And having meta language that is significantly different from the core language really sucks for ergonomics. I never understood why, for example, a lot of dependent types literature invents a meta language for expressing constraints rather than using the core language for it. Syntactic familiarity matters.

Also consider C++ templates vs Zig comptime, and how C++ is evolving to bring its meta language closer to the core language via constexpr and consteval.

So the more viable path is to judiciously integrate formal concepts into "normal" programming languages and have LLMs generate proof artefacts in a search process that is guided by proof systems.

Sure, which is why right to repair laws are so important.

I understand this discussion as being about how society should deal with it, not how you could try to make the argument internal to a company.

> But for safety reasons, since the software controls when the toaster pops, you decide to check at boot time that the software hasn’t been modified.

As arguments go, this is a pretty weak one considering how obvious the solution is: Make the manufacturer not be liable for what happens when you operate the device with unauthorized software.

Plus, Rust has a more expressive type system.

I like loose type systems for some quick scripting, but I started to adopt Rust for many of my personal projects because I find it's so much easier to get back into a project after a year when there are good type system guard rails.

Discrete GPUs still have an advantage in memory bandwidth. Though this might push platforms like laptops towards higher bandwidths, which would be nice.

We're all groping around in the dark here, but something that could have happened is a tokenizer artifact.

The vocabularies I've seen tend to prefer tokens that start with a space. It feels somewhat plausible to me that an LLM sampling would "accidentally" pick the " Hello" token over the "Hello" token, leading to D:\ Hello in the command. And then that gets parsed as deleting the drive.

I've seen similar issues in GitHub Copilot where it tried to generate field accessors and ended up producing an unidiomatic "base.foo. bar" with an extra space in there.

> LLMs are humanity's "first contact" with non-animal intelligence.

I'd say corporations are also a form of non-animal intelligence, so it's not exactly first contact. In some ways, LLMs are less alien, in other ways they're more.

There is an alignment problem for both of them. Perhaps the lesson to be drawn from the older alien intelligence is that the most impactful aspect of alignment is how the AI's benefits for humanity are distributed and how they impact politics.

Sample size is way too small.

Unless he was able to sample with temperature 0 (and get fully deterministic results both times), this can just be random chance. And experience as SWE doesn't imply experience with statistics and experiment design.

The counter-devil's advocate[0] is that consuming CSAM, whether real or not, normalizes the behavior and makes it more likely for susceptible people to actually act on those urges in real life. Kind of like how dangerous behaviors like choking seem to be induced by trends in porn.

[0] Considering how CSAM is abused to advocate against civil liberties, I'd say there are devils on both sides of this argument!

You're probably right, but that doesn't mean GP is wrong, just that they need to state their thesis more carefully. There's more science to be done there, but there's no reason to believe that new fundamental laws of physics are required to explain the brain.

If you continue that thought experiment, you'd very quickly run into the issue that the way the 386 interfaces memory is hopelessly primitive and not a good match for running 1000s of cores in parallel.

A large reason why out of order speculative execution is needed for performance is to deal with the memory latencies that appear in such a system.

It's important to remember that people are not some homogenous mass.

There were certainly some rich and powerful people who did intend to lower everyone (else) to a shit position.

That's what stuff like anti-union policies are about. Sure, they drape it up in nice rhetoric, but ultimately the intent is to reduce the power (and thereby the material outcomes) of the common the worker.

But of course there were also other developments that had the effect more unintentionally.

Let those large retailers put pressure on their suppliers. Prices haven't exactly been stable recently. I really don't think it matters, but if it did (as you claim) then surely some downward pressure is a good thing.

Which is ironic, considering that the 1950s were long before the internet boom. The internet didn't even exist yet, let alone dial-up modems.

FWIW, it's part of the plot of Andreas Eschbach's Lord of all Things. (IMHO one of his weaker novels, but weirdly one of the few that were translated into English)

Aside from the fact that such calculations aren't necessarily applicable anyway, it is incorrect because they would most likely have continued to use and have to pay for Evernote for more than just the one year.