dr_zoidberg

MS Paint wanted to add the "this is AI" tag on a picture they just resized. OP didn't like that, so they went and downloaded Paint.Net to avoid having to deal with (MS) Paint shenanigans.

Your suspicion could have easily been cleared by reading the paper.

If you're short on time: the paper reads a bit dry, but falls in the norm for academic writing. The github repo shows work over months on 2024 (leading up to the release of 3.13) and some rush on Dec 2025 to Jan 2026, probably to wrap things up on the release of this paper. All commits on the repo are from the author, but I didn't look through the code to inspect if there was some Copilot intervention.

[0] https://github.com/Joseda8/profiler

If we go by Microsofts 2020 account of 1 billion devices running Windows 10 [0], and assume all those are running some kind of electron app (or multiple?) you easily get your gigawatt by just saving 1 watt across each device (on average). I suspect you'd probably go higher than 1 gigawatt, but I'm not sure as far as making another order of magnitude. I also think the noisy fan on my notebook begs to differ and maybe the 10 GW mark could be doable...

[0] https://news.microsoft.com/apac/2020/03/17/windows-10-poweri...

Not sure about that, SSDs historically have followed base-2 sizes (think of it as a legacy from their memory-based origins). What does happen in SSDs is that you have overprovisioned models that hide a few % of their total size, so instead of a 128GB SSD you get a 120GB one, with 8GB "hidden" from you that the SSD uses to handle wear leveling and garbage collection algorithms to keep it performing nicely for a longer period of time.

It's at the end, in the "What are the standards units?" section.

Yes, the thought crossed my mind too... But then I tried a private window and it opened, so maybe the other suggestion that the cookies are very long lived is right.

A little bit off topic: but I couldn't even start to read the article because "I reached my article limit" out of I site I never visited before... What are they using to determine how many articles I've read?

Opening in a private window solved the issue, however I'm pretty sure I don't regularly read anything on this site (maybe never was an overstatement?).

In the ~30 years I've used computers, they've become ~1,000,000 times faster. My daily experience with computers doesn't show it. There's someone out there who took the time to measure UI latency and has shown that, no only isn't it faster, it's actually slowed down. And yet, our hardware is 1,000,000 times faster...

Edit: this is the latency project I was thinking about https://danluu.com/input-lag/

The abstract of OPs link mentions "Processing-Using-DRAM (PUD)" as exactly that, using off the shelf components. I do wonder how they achieve that, I guess fiddling with the controller in ways that are not standard but get the job (processing data in memory) done.

Edit: Oh and cpldcpu linked the ComputeDRAM paper that explains how to do it with off the shelf parts.

No, I just went to search if the topic is mentioned in guidelines (which it is, multiple times). I'd then expect a (good) expert to pick on those breadcrumbs and search on how to do that (if they don't have the skills already). If I were working on a computer, I'd try to find IOCs that point to an infection (or lack of evidence for it).

If there's a memory dump to work on, a more in-depth analysis can be done with Volatility on running processes, but it usually falls back on the expert having good skills on that kind of search (malfind tends to drop a lot of false positives).

But at least the guides gave a baseline/starting point that seems to be better than what was described. It's very difficult to prove a negative, so I'd also be careful with the wording, eg: "evidence of a malware infection was not found with these methods" instead of "there's no malware here".

The lack of standards falls on the acting part. I ran a quick search and found that SWGDE best practices guides and documents do consider the case for the presence of malware on the digital evidence sources on many different scenarios [1]. Having an "expert" who is unaware of these guides is another story.

[1] https://www.swgde.org/?swp_form%5Bform_id%5D=1&swps=malware

That's interesting. A project at work is affected by Windows slow open() calls (wrt to Linux/Mac) but we haven't found a strong solution rather than "avoid open() as much as you can".

As of now 9%. I thought hitting the HN front page could have a much larger impact on this, but it seems that's about it this time.

When I moved some projects from OpenCV 3 to 4 I got a nice speed up pretty much everywhere, some things no speed up at all and some others pretty big. I can't really remember the numbers, but at the moment it was a global 10 to 20% perf improvement just on updating a library.

Might want to check that. Also 4.something got SIFT as part of OpenCV (instead of living in the contrib module) because the patent expired and you can now use it for free.

As for blowing up with NN packages and such... I don't really use those parts, but if the NN module had easier support to run networks trained on popular frameworks I might've used it. Disclaimer: it's been quite a while since I last tried to use those parts, so maybe now the latest version has fantastic support and I'm talking nonsense.

Indeed, it's their trademark style. It has also worked great for them.

I thought I might have missed a rename. Microsoft tends to rename things at random times.

I ran a few tests in Windows Terminal. The bomb emoji got width 2, while the motorboat got width 1 and correct aspect ratio, though I didn't quite get to see it properly until I zoomed in like 5x. The family was rendered all as one emoji, cells wide, but left 4 blank cells before it.

So it was a bit better than the authors tests, but there's still room for improvement.

Is that Windows Terminal, or have they renamed it yet again?

> It means LK99 is about as good as the other materials, at minimum. However, its chemical composition is different than any of the other materials on the chart (i.e. it is lead based)

While what you say is correct, let's not forget that lead _is_ an elemental superconductor. With a Tc of 7K, it's only bested by niobium (9K) and diamond (11K) as an elemental superconductor.

Yes, cuprates and ceramics were ruling high-temperature SC so far, but it's not like lead was entirely unexpected in the superconducting world.

> But then why “20 years”? Would performance at the level of a 30-year professional be considered failure?

They're keeping the "30 years of experience" agent for the 2.0 version /s.

This is a good analysis, and I'm glad that there are some people like you giving these things a good deal of thought.

It's also 3.6 times faster, in a world where hardware updates get you 20-ish % improvements per generation.

Today the upper range of RAM in devices is in the hundred-GB. 10 years ago it was about 16GB or so, and making a (probably bad) interpolation I think it wouldn't be crazy to have the upper range in the TB mark in 10 years time. We could get there faster too, for specific use cases (compiling, rendering, etc).

On the other hand, yes, everyday use (web, mail, video/media consumption) don't require today much more than 8 or 16GB of RAM. If we go a bit creative, a PC with Linux can run very smoothly on 4GB alone, and surely someone here can point to their one anecdote of a machine with 2 GB or even 1 GB sporting a nice desktop environment, or 128mb CLI-only machine.

Edit: also memory has to improve its bandwidth and data transfer rates to keep up with faster processors, so it could also improve over 10 years time without much focus on storage capacity. Or maybe they focus on latency instead, or a mix of all three. Point is that it's not a single metric to improve.

I'm curious about your view on ONNX. At work we did a few prototypes and it seemed to work well enough for our use cases, and we're moving to it. What is it that we haven't seen yet that gave you trouble?

Admittedly we're on a reasonably easy situation: we just have to deploy models (some from scikit-learn, some from Keras, some from PyTorch) to various users who mainly run a specific version of python under Windows and Linux, with CPU and GPU support.

No need to quote you, but GvR himself [0]. That's the Faster CPython initiative, which has been ongoing for about 2 or 3 years I think. 3.11 got some nice speedups from it, and 3.12 is on its way to get more. [1]

[0] https://github.com/faster-cpython/ideas/blob/main/FasterCPyt...

[1] https://github.com/faster-cpython/ideas

One other thing. The article states pythons performance at keys-per-second, then all the other numbers are given in keys-per-minute. That means that pythons number looks really small in comparison, but if we take it (500 kps) and multiply it by 60, we get 30k keys-per-minute, or about 50% faster than the first CUDA baseline.

Now I'm quite fond of python, but largely I read this as the CUDA implementation having quite some room for improvement. Having almost no CUDA experience of my own, that is just a hunch, which I'm glad rfoo supports with (surely) a lot more experience than me.

I've been using a Zen2 notebook for the past few years, and was honestly surprised by the processors performance for the first... 3 or 4 months. Then typical updates happened. And some quirks got in the way too. Like it having a decent iGPU that I can't use because the BIOS will only let me pick one of them (down to it not having a mux to pick one or the other? Not sure what's going on there, just that I can't do it). In general it's a great machine, but there's a host of little details that make the experience a bit worse than it should be.

Per my usual update schedule, I'm looking down the line to at least a Zen5(+?) upgrade in a few years time, so I hope they improve this kind of things in the future. However that's entirely up to OEMs deciding to make a good product, and a bit out of AMDs grasp.

And that's why so far AMDs mobile processors have been monolithic and not chiplet-based. That is supposed to change with Zen 4's Dragon Range, however most of the mobile lineup will still be monolithic and these high-power/high-performance processors should go exclusively to "gaming" notebooks.

Indeed, in this case ESRI is talking about deep-learning based single image superresolution. With multiple images, even something as simple as shift-and-add can recover details (and lower noise in the picture), but you do require having multiple images. Video being a sequence of images, with large or small movement in between frames, can be an ideal source for robust SR algorithms. To complicate things further, there are deep-learing based multi-image SR algorithms too.

Yes, the Galactica LLM by Meta. Though LeCun isn't an author in the paper, he is "Chief AI Scientist for Facebook AI Research (FAIR)"[0], and he was quite angry about closing the Galactica demo[1].

[0] https://ai.facebook.com/people/yann-lecun/

[1] https://twitter.com/ylecun/status/1593293058174500865

That comes from mimicking Matlabs plotting API, as a way to move people from Matlab into scientific python. "It's a feature not a bug", a bit taken to an extreme. But it was successful.