phdelightful

Thank you, I learned a few things!

I didn't check thoroughly, but the first one I happened to grep out was not on that list:

"Mozilla/5.0 AppleWebKit/537.36 (KHTML, like Gecko; compatible; Amazonbot/0.1; +https://developer.amazon.com/support/amazonbot) Chrome/119.0.6045.214 Safari/537.36"

"x-forwarded-for":"44.210.204.255" "x-real-ip":"44.210.204.255"

This is a bit outside my area of expertise, so I don't know how reliable these x-forwarded-for and x-real-ip are.

I just put Anubis in front of my self-hosted forge this morning because AmazonBot had helped itself to 750 GiB (!) of traffic to my public repos this month!

At least, it claimed to be AmazonBot…

Parent's article says

Starting from version 6.6 of the Linux kernel, [CFS] was replaced by the EEVDF scheduler.[citation needed]

It’s <= a Radeon 7600 GPU (28 CUs RDNA3 vs 32), so I’m not sure I’d have advertised it as a 4k60 machine. Then again I’m not a marketer so what do I know. 4k60 is a flexible target with FSR I suppose.

What coordinate in the space is furthest from any named color? It looks like there are some relatively large voids in the blue/purple boundary area but it’s hard to say.

I think I’d buy something with Strix Halo or Strix Point if there was official ROCm support. As of 6.4.1 from earlier this month there’s still not, as I understand it. I’d be delighted to be corrected on this matter.

The article goes on to say what the author thinks is bad about this:

We’re not raising emotionally intelligent kids. We’re raising kids to navigate human unpredictability as if it’s a design flaw. Because when you grow up with a machine that always gets you, messy human behavior feels broken. We’re not preparing kids to handle people.

I don’t think there’s anything wrong with escaping into fantasy in the right time and place, but young kids (and even well-adjusted adults) can have problems self-moderating and letting fantasy substitute for engaging with reality.

I compiled it for Ampere and counted 6834 actual F32 operations in the SASS after optimizations. I only counted FFMA, FADD, FMUL, FMNMX, and MUFU.RSQ after eyeballing the SASS code, so there might even be more. It's possible the FMNMX doesn't actually take a FLOP since you can do f32 max as an integer operation, and perhaps MUFU.RSQ doesn't either, but even if you only count FFMA, FADD, and FMUL there are still 3685 ops.

  nvcc -arch=sm_86 prospero.cu -o prospero
  cuobjdump -sass prospero | grep -E 'FFMA|FADD|FMUL|FMNMX|MUFU\.RSQ' | wc -l

I basically have an even simpler version of something like this for my own personal use too. I found it pretty easy to write in Go and my area of expertise is decidedly not web frontend/backend. I’d recommend it as a fun little project if you’re looking for something to do.

For mine, I paste in a video or playlist URL and it downloads the video and creates a lower resolution transcoded version suitable for streaming to my phone. It also extracts an audio-only version in case that’s more appropriate.

The Ars Technica article:

https://arstechnica.com/tech-policy/2025/02/youtube-briefly-...

(PS: Ars Technica is a bit sluggish for me this evening. Not sure why.)

Since you asked for “all the feedback,” there’s a typo on your landing page:

“The Meha API utilizes it's home-grown” -> “its”

Also, I got a relay access denied error when I tried to email you at info@meha.ai

My reading of OP is that it’s less about whether zopfli is technically the best way to achieve a 5% reduction in package size, and more about how that relatively simple proposal interacted with the NPM committee. Do you think something like this would fare better or differently for some reason?

Yeah, we really just try to come up with very loose bounds since the analysis is hard. Even so, it does occasionally stop us from getting things way way wrong.

I have worked on a performance-portable math library. We implement BLAS, sparse matrix operations, a variety of solvers, some ODE stuff, and various utilities for a variety of serial and parallel execution modes on x86, ARM, and the three major GPU vendors.

The simplest and highest-impact tests are all the edge cases - if an input matrix/vector/scalar is 0/1/-1/NaN, that usually tells you a lot about what the outputs should be.

It can be difficult to determine sensible numerical limit for error in the algorithms. The simplest example is a dot product - summing floats is not associative, so doing it in parallel is not bitwise the same as serial. For dot in particular it's relatively easy to come up with an error bound, but for anything more complicated it takes a particular expertise that is not always available. This has been a work in progress, and sometimes (usually) we just picked a magic tolerance out of thin air that seems to work.

Solvers are tested using analytical solutions and by inverting them, e.g. if we're solving Ax = y, for x, then Ax should come out "close" to the original y (see error tolerance discussion above).

One of the most surprising things to me is that the suite has identified many bugs in vendor math libraries (OpenBLAS, MKL, cuSparse, rocSparse, etc.) - a major component of what we do is wrap up these vendor libraries in a common interface so our users don't have to do any work when they switch supercomputers, so in practice we test them all pretty thoroughly as well. Maybe I can let OpenBLAS off the hook due to the wide variety of systems they support, but I expected the other vendors would do a better job since they're better-resourced.

For this reason we find regression tests to be useful as well.

Here’s a blog that breaks down how large different pieces of CUDA are:

https://carlpearson.net/post/20231023-cuda-releases/

It’s been a long time since I worked on FPGAs, but it sounds like FPGAs! What do you see as the main differences?

Your impression is correct. Peer review would never catch this. Peer review basically assumes the counter party is operating in good faith, and as a result a thorough peer review basically is the following:

* is the treatment of existing work semi-thorough (even experts don’t know everything) and fair?

* are the claims novel w.r.t the existing work? If not, provide a reference to someone who has already done it.

* can you understand the experiments?

* do the experiments and their results lead to the conclusions claimed as novel?

* does the writing inhibit understanding of the technical content?

No peer review I have ever seen or done would catch anything but the most egregious bug of this nature.

DOE HPC people are not having this problem with AMD GPUs: here’s a paper that reports 1.36 TB/s on one GCD of an MI250x (theoretical is 1.6TB/s).

https://dl.acm.org/doi/pdf/10.1145/3624062.3624203

Table 6

One of my colleague's Ph.D. thesis was on how to achieve high-performance CPU implementations for bulk-synchronous programming models ("GPU programming")

http://impact.crhc.illinois.edu/shared/Thesis/dissertation-h...

It’s far from what I do now, but as an undergrad I built an automatic guitar tuner. It had a PIC32 microcontroller that read the sound waves on GPIO pins and did some cross-correlations to figure out the frequency (sort of a poor man’s FFT, but faster since it was less general). It used an FPGA to drive a stepper motor to turn the guitar pins towards the correct frequency. Code was all C and Verilog.

In my line of US government coding (no relationship with this project), NDA is orthogonal to security. NDA is used to protect confidential vendor information. For example: that they have a contract with the government at all (in the case of a stealth startup), specific technical capabilities they don’t want broadcast to their competitors, the size of the contract vs. committed resources, etc

Interpolate: to alter or corrupt by inserting foreign matter; but I have only seen it used this way for text

From my perspective, it seems two main things limit LLM adoption in my area of the Department of Energy. I'm not in management, so I don't have any particular insight in the procurement process.

1. Information sensitivity. Even ignoring classified information, there are quite a few things we can't even put into a Google search. It's definitely a no-go for this to end up in a training dataset.

2. "Hallucinations"

Making LLM available through some infrastructure that is already approved for sensitive information will definitely help with the first point, and allow us to experiment with more areas where it might be helpful. I presume this would come along with guarantees about the interactions not being used for training.

It might be even better if some company would sell an appliance we could install on-prem with similar non-training guarantees. Then we could leverage these new tools for very sensitive information, which could be a great help.

I achieved my quickest, simplest, most pleasant customer service interaction yet with an airline by writing a letter explaining what they did (with transaction IDs, confirmation numbers, etc), and what they should have done. The situation was not entirely simple, and I was not looking forward to explaining it verbally to a rep.

They sent me back a simple letter a couple weeks later saying they'd resolved the issue the way I requested and provided me the updated documentation.

I'll definitely consider this approach next time something like this happens.

Framed another way, what you're describing here is that using ChatGPT is interesting enough to trigger engagement from social media users who are otherwise mostly inactive. If that is indeed the case, imo that is a pretty strong indicator of appeal to a wide audience.

I understood the comment to mean the posters are not posting about ChatGPT, but using ChatGPT to generate vapid (?, or inauthentic?) posts to farm engagement where there would otherwise be silence.

TFA says that their app loads faster for this reason with the new implementation.

If you can tolerate a .gov email address, there are high-performance computing, edge computing, and AI groups at DOE labs. My impression is that they're always starved for talent in these areas due to stiff competition from industry and lifestyle restrictions that come with a security clearance (being a US citizen, reduced international travel, no drugs other than alcohol and nicotine, your work may require on-site). Defense labs are probably similar.

People who feel strongly that current procurement strategies are the wrong way to go about things should devote as much toil and treasure as they can to seeing their preferred solution materialize. In the mean time, I hope they will allow others to work within our existing system to keep our adversaries from replacing the US with something much worse for the entire world.

I basically agree. I don't think the political mechanisms exist in the short-to-intermediate term to get salaries in government labs high enough. As you point out, the extent to which "civilians" will be willing to participate in these lines of work is an open question (said without judgement, just as an observation). Presumably folks who have opted-in to work at defense contractors will be more willing, but do they have the ability to attract a sufficient amount of talent in this area?