The Ampere cpu in that post has a "lack of single core CPU speed" despite ripping through compilation tasks with its 80 cores. But the DGX Spark's CPU has fairly decent single core speed and is more like Apple Silicon in this respect.
It's an Orico 9948C3 with four Seagate Barracuda 24TB drives. They were on sale last year [1].
Unfortunately, the enclosure doesn't work super well on Linux.
There is a weird bug where the drives don't enumerate when I boot up my computer. This happens on both my x86_64 AMD machine running Linux, and on the DGX Spark. The solution is... simply power cycle the enclosure a couple of times by toggling the power button on it and then it works. Once all four drives show up in lsblk, I can `sudo zfs import ...` manually. This is really gross and annoying. Replacing the USB cable, flipping the USB-C cable 180 degrees, hot plugging it, etc, all didn't work, both on the DGX Spark and the other Linux machine. I've also read reports of it being unstable in UAS mode on Linux but I haven't found a big difference in stability between enabling UAS or falling back to usb-storage.
Once it starts up correctly though, the drives are fast. I store my huge amount of 100 megapixel photos on it.
The Seagate Barracudas are helium-filled HAMR/CMR drives and are apparently rebranded/binned Exos drives. They aren't rated for 24/7 use but then neither are the refurbished Exos drives.
Yeah... that's been my past experience with USB docks, at least any with more than one slot. I've never had great luck with them. "Can recover from power outage without being touched" is a key requirement for my NAS so I'll give that one a pass and stick with my "SATA drives directly attached to a SAS controller" strategy for now. Thanks for the reply.
As for the 24/7 use, yeah so be it. The I in RAID stands for Inexpensive. If they fail after 10 years at 24/7, so be it. I have drive level redundancy and frequent offsite backups of anything critical.
It's absolutely the stupidest solution, but SwitchBot Bot is a home-automation button pusher. They are quite cheap on ebay (especially compared to the price of hard drives) but still a face-meltingly stupid problem to have. https://us.switch-bot.com/products/switchbot-bot
I have quite a mix of 2x and 4x drive enclosures, and generally I don't notice anything particularly odd about them? Some enclosures have degraded smart or power management capabilities though. I really wish I'd noticed at the time & sent those ones back! What a waste! My favorite are some dual slot HighPoint toaster-style ones that have independent USB connections per drive, which rocks, but these are so rare & I almost never take advantage of it (replug on drive into a different mini-pc).
The DGX Spark hits a sweet spot for me where it can simultaneously work for general computing and run local LLM inference fast enough for some hobbyist dabbling. The RTX 6000 Pro Blackwell is more than twice the price (it has increased quite a bit recently from $8000 to $11600), not to mention the "rest of the PC" needed to get it working, so it's not really a fair comparison. Compared to other 128 GB unified memory devices like the Mac Studio and the Strix Halo, the DGX Spark fairly priced in my opinion.
I have a AMD Strix Halo box I use for similar dabbling. It definitely wasn't an "out of the box" experience, fiddling around with kernel versions and ROCm installs. These days I mostly wind up using the Vulkan build of llama.cpp for inference.
Many other distros almost work out of the box (as in they boot and run without any modifications). The custom kernel patches you mentioned address mainly non-critical bugs such as a bug where the Realtek r8127 stops working after a reboot (but it works if you turn it off and on again) [1] [2]. I'd consider it in a way better state than trying to run other Linux distros on certain device tree-based devices like, say, Qualcomm Snapdragon machines. The regular NVIDIA drivers with the open source kernel modules work just fine. Talos Linux supports DGX Spark since version 1.12 [3]. I also know of people using Fedora and nixOS successfully.
iirc the bug that the R8127 patch was addressing was fixed in the R8169 driver upstream in kernel 6.18 and later. Not super sure what other patches there are and what the status of those are...
> Every single one is a pelican, on a bicycle, first try. When every student gets an A, the exam has stopped grading.
Numerous pelicans and their bikes are clearly horribly malformed. In fact none of the bike frames are correct. Fable and Opus come close, but the top of the diamond is disconnected in Fable's case and the head tube is misaligned with the front fork in Opus's case.
And of course, as the parent post shows, labs don't actually seem to be training on the pelican bike case.
Agreed. And more; the Macbooks are pretty much the same - some are god approximations, some are terrible, all of them are recognisably a MacBook. And if you start using it they can train on it.
The problem isn't the test, its that is a public test.
Simon has previously said he has a list of secret prompts (at least one of which he "burned" as a demonstration a while ago). That's what makes it a good test - his commentary on the public test is something of a proxy for non-public tests. This makes it a good benchmark.
I feel like getting LLMs to spit out an SVG is akin to getting a human artist to draw something by just reciting a list of coordinates. It's insanely hard and unnatural.
Image generation models nowadays can easily generate a photorealistic pelican riding a bicycle, where the bicycle has perfect structure. But it is, of course, only a raster image.
It seems that we're missing a kind of step to decompose an image into a list of instructions (say, SVG paths, or even brush strokes with a real brush) to reproduce it properly. Doing so would probably need a true understanding of the structure of the scene, which is something that AI still struggles with to this day.
I know. I'm just thinking about how to make AI create SVGs better... in theory, a sufficiently smart AI could "generate an image in its head", think about it, and then output the SVG paths to produce said image. Intuitively that would be somewhat closer to how human artists convert artistic visions into a sequence of arm movements while holding a brush (obviously, humans don't hold a fully formed, photorealistic image in the head while drawing, but rather vague concepts, but still).
I don’t quite agree. Good human artist can visualize in their mind how to draw a picture, i think. Which i think is no different than LLM doing SVG drawing in their “head”. Anthropic’s recent post call this head-space “workspace”.
It just might feel foreign to human who does not have a SVG trained head-space.
Sorry, my comment was confusingly worded, but I am precisely thinking about the head-space thing you are talking about. My point is: the fact that image generation models can generate a perfect pelican riding on a topologically correct and highly detailed bike frame [1], suggests that AI does have the capability to correctly understand it internally. But something is disconnected when we just let an LLM directly output the SVG, and all the nuanced understanding is lost.
As such, modern LLMs still kind of suck at generating pelican bike SVGs with obvious errors:
* some omitted the bottom of the diamond which connects from the pedals to the rear wheel
* some added an extra connection from the pedals to the front wheel, making it impossible to steer
* none could align the head tube with the fork
* none added a correct offset to the fork
* none could generate the chain properly in a way that attaches to the two sprockets correctly
... whereas these errors do not appear in the raster image. (To be fair, the raster image has other weirdnesses, like the bird having arms and only one leg)
If we could harness the sort of internal thinking that must have happened when it generated the highly consistent raster image, but make it output SVG instead, we'd get much better SVGs. So this would be indeed "visualize in their mind how to draw a picture" before outputting the SVG.
Image models that support text output like Image2, or general text models that can read images like Claude can vectorize raster images. But they aren't very good at it, doing it manually in Inkscape still produces better quality even when done by non-artists.
On an unrelated note: I found this map where you can view the changes in political geography over time via scroll at the bottom. It's my goto with historical posts.
I totally understand the point of view from maintainers. Review fatigue of low quality slop is a legitimate issue.
The worst ones are fully autonomous AI agents looking for open source projects and adding random pull requests.
But in some cases, I find a legit bug that needs fixing. For example, I want to get a particular program working in Wine/FEX on aarch64 [1], or I find a 12 second hang in Darktable [2]. The problem is that, as a software engineer working in a totally different discipline, I have no knowledge of the low level C code to fully understand what the problem even is, or how to fix it. All I want to do is to fix the issue and help other people avoid running into the same issue. Right now, on my machine, I maintain a set of custom patches to get everything working. But I am too dumb and ignorant to figure out how to create the fix by hand, so I can't submit a pull request (or when I do, I feel really bad about it. I honestly feel like a horrible person, e.g. when a project added a "No AI" policy soon after I submitted some AI-generated PRs [3]). Going forward, I feel like this sort of scenario is going to be way more common.
Then say all of that in the issue. Say you have a real problem. Say you tried using AI. Add the human element by communicating. I dont think there is a real problem there.
People just want to know you put in the effort, and that you didnt just prompt an AI and hand over completely unchecked slop.
For my personal website I have a lot of photography-oriented blog posts [1], but I have special code to strip out GPS info from the location if it's close to my home [2].
EDIT: my vibe-coding slop agent put my home GPS lat long in the example config in the README lol. Please don't rob my house; I'll go run git-filter-repo later.
Yeah, if I take a dense grid of photos near my house, it would reveal a 500 m circle. But in practice I don't take _that_ many photos in the neighborhood. Also, the circle isn't perfectly centered on my home.
reply