bmenrigh

A pretty reasonable question. I guess it could be my ISP filtering.

Or rather could have been, since the issue seems like it was transient. I'm now getting resolutions.

Very strange, I wish I'd recorded what was going on better while it was broken.

I'm in the US and I'm either getting NXDOMAIN or NOERROR (with no A record answer) back from every big nameserver I try (my ISP, Google, Cloudflare, etc.).

But a dig +trace archive.ph (which recurses all the way to the root locally) resolves it fine.

Is there some US-mandated DNS filtering I don't know about?

The 2n choose n solution isn't at all intuitive to me but thinking about it in terms of 40 steps, 20 of them rightward and 20 of them downward an then looking at all distinct permutations of these 40 steps as (40!) / ((20!)^2) is intuitive to me. Then it becomes obvious that since 20 is half of 40, k and n - k are the same number (20), which coincides with the binomial coefficient n! / k!(n - k)!. But this seems like a lucky coincidence in 2 dimensions and if you extended the problem into 3D you'd do better thinking about permutations.

Oh I didn’t mean at all charging them. I mean licensing in the sense of granting rights for the purpose of training. Probably most labs would be fine adding the language to the training for free as long as the dataset quality is high and it improves the results. But yes, pay them if that’s what it takes for them to use it.

I think at this point what the Netherlands, and any other country that wants a good model in their language should do, is gather up every piece of text ever written in that language and license it to the big AI labs/companies for training. I'm sure there are vast libraries of books and other text that haven't been digitized and aren't a priority for the big labs.

I was just looking at Unicorn last week because it's used by unipacker to do automated unpacking of binaries. I built a "toolbox" for gpt-5.5 to do semi-automated malware and exploit reverse engineering and unipacker is sometimes useful for that purpose.

Mini Micro seems to be built on Unity. The MiniScript portion of it is open source https://github.com/JoeStrout/miniscript but the version packaged for use by Unity costs some money. I can't tell if the people behind MiniScript are the same people behind the Mini Micro.

Every C program I've had codex write ended up costing me more time than had I just done it from the start myself. Whereas almost every Python program it's written for me saved me time, even including the time I spent cleaning it up.

I chalk this up to primary two reasons. First, I cared a lot more about the implementation details of the C program than I did the Python one, and second, it's just better at simple stand-alone python programs than it is at C programs.

The criteria I know use is "do I care about the implementation details of this?". If I do (because for example it's going to be long-term code that I need to maintain) then the agent likely isn't worth it. But if I don't, there are huge efficiency gains to be had using the agent.

The black box analysis needed to decode this is incredibly hard but also incredibly fun and rewarding to pull off. Very impressive work.

Ah I was a few years later on the TI-86. Around 1999.

Between this, and SNES emulation (searching memory for values and adjusting things to see how they'd affect the game), I was destined for computer science.

Which is exactly why I mentioned rejection of zero, negative numbers, etc.

You can reject them, but doing so just throws away useful tools without gaining anything in return.

Rejecting infinity is a purely philosophical stance that doesn’t teach us anything about reality.

There is a big difference between “infinity doesn’t exist” and “infinity doesn’t exist physically”.

I should also add that the resolution of zeno’s paradox in the form of calculus where and infinite set of steps can occur in a finite time (or infinite set of distance can span a finite total distance) is conceptually very simple and useful. Rejecting it as unphysical, or saying it must imply time or space come in discrete chunks, is not contributing to an understanding of reality unless the rejection also comes with a set of testable (in principle) predictions.

The article doesn’t really tell us what is gained by rejecting infinity.

And in general, why not also reject zero, negative numbers, irrational numbers, complex numbers, uncomputable numbers, etc.?

Seems like an article about quacks that can’t even agree on what the bounds and rules of their quackery are.

Reminds me of around 2002 when MS slapped “.Net” onto everything.

Yeah but to the commenter I was replying to, I don't think it was clear that detail was relevant to the benchmark numbers they were quoting.

No, the gains here aren't very dramatic when compared properly (against fsync), and have nothing to do with AI help. The gains come down to Linux kernel support for certain synchronization primitives like the Mutex on Windows, such that there is a more direct mapping of what a Windows binary expects to what the Linux kernel provides. See https://docs.kernel.org/userspace-api/ntsync.html for the kernel support that makes this possible.

Those benchmark numbers are slightly misleading, as they are a comparison of Wine+ntsync against Wine+nothing. There has been a somewhat fast "fsync" library built around Linux's futex and the gains over Wine+fsync are modest (just a few % in most cases).

That said, Wine+ntsync is still a win, just not a 8x improvement like the Dirt 3 benchmark suggests.

(And it case it's not clear, ntsync is https://docs.kernel.org/userspace-api/ntsync.html, which is a driver for Linux that offers syncronization primitives (mutex, semaphore, events) that more closely match the semantics of the Windows primitives. It's easier to do a direct implementation in Wine to support code compiled for Windows that expects to be talking to an NT kernel.)

I don’t fault Quanta (or 3b1b) for being the way they are. Each is serving their goal audience pretty well.

My compliant is only that there should be a dozen more just like them, each competing with each other for the best, most engaging math and science content. This would allow for more a broader audience skillevel to be reached.

As it stands, we’re lucky even to have Quanta and 3b1b.

I think there is hope though, quite a few new-ish creators on YouTube are following in Grant’s footsteps and producing very technically detailed and informative content at similar quality levels.

Modern x86_64 has supported multiple page sizes for a long time. I'm on commodity Zen 5 hardware (9900X) with 128 GiB of RAM. Linux will still use a base page size of 4kb but also supports both 2 MiB and 1 GiB huge pages. You can pass something like `default_hugepagesz=2M hugepagesz=1G hugepages=16` to your kernel on boot to use 2 MiB pages but reserve 16 1 GiB pages for later use.

The nice thing about mimalloc is that there are a ton of configurable knobs available via env vars. I'm able to hand those 16 1 GiB pages to the program at launch via `MIMALLOC_RESERVE_HUGE_OS_PAGES=16`.

EDIT: after re-reading your comment a few times, I apologize if you already knew this (which it sounds like you did).

I recently started using Microsoft's mimalloc (via an LD_PRELOAD) to better use huge (1 GB) pages in a memory intensive program. The performance gains are significant (around 20%). It feels rather strange using an open source MS library for performance on my Linux system.

There needs to be more competition in the malloc space. Between various huge page sizes and transparent huge pages, there are a lot of gains to be had over what you get from a default GNU libc.

Yeah I don’t get it. Shifts and rolls are among the simplest of all instructions to implement because they can be done with just wires, zero gates. Hard to imagine a justification for leaving them out.

My main skepticism here is whether the theorems have been properly replicated in the proof. Verifying that the proof really captured the mathematical statement seems like a manual, human process, and quite hard to repeat reliably across proofs.

Yeah I learning about LPCAMM2 memory was far more interesting than the repairability score.

Oh completely. But my perspective is that we all should individually punish clickbait by not clicking. More broadly, we should strive to keep HN full of quality tech content rather than clickbait.

"better", "more accessible"? What the hell are you talking about? Clickbait doesn't make anything better or more accessible.

Instead, it makes it impossible to pre-select for interesting information. Instead of telling you what something is about, it tells you how you should feel about it. That's not improving accessibility.

Can we stop it with "and the results are terrifying", "and you won't believe what I found", "the <x> situation is insane", etc.? The over-hyping of low quality, low effort content is making it hard to find actually interesting or informative things.

I love the AVX512 support in Zen 5 but the lack of Valgrind support for many of the AVX512 instructions frustrates me almost daily. I have to maintain a separate environment for compiling and testing because of it.

I should add that I have a group of friends that only chat on "C", for C in:

* SMS * Apple iMessage * Email * IRC * Facebook Messenger * Telegram * Slack * Webex Teams * Discord * Twitter (DMs) * Signal * Whatsapp * A particular PHPBB web forum

Instead of being on top of all of these, I mostly neglect all of them, and then friends complain that they haven't been able to get a hold of me for 6 months.

XMPP was my own "solution" to this problem, which nobody else used.

I ran my own XMPP server for about 15 years. Then 10(?) years ago Google's GChat migrated away from XMPP. I never had another XMPP conversation with anyone from that point on. I finally turned the XMMP daemon off off about 2 years ago to reduce my attack surface.

Often "nothing is changing" ends up being more literal than the founders realize or intend. Acquisitions by big companies tend to slow the development to a crawl as development bureaucracy takes over. When a great product is practically frozen in time it stops being great in 5-15 years as the rest of the world passes them by.