hedgehog

15 ms·

241.
by hedgehog 7mo ago
I'll take a reasonable path over no path and just hoping VirtualGL or something will be enough and forgoing color management entirely. I understand that some use cases work better or only in X, but I also see the roadmap for Wayland an
242.
by hedgehog 7mo ago
Remote access on X11 is a mess and I won't miss it, at least on Wayland everyone is funneled through EGL or Vulkan and there's a reasonable path to layering remote access on top of that.
243.
by hedgehog 7mo ago
Bunnie did a really good talk a couple months ago that has more of the background beyond what's in the blog post: https://www.youtube.com/watch?v=H5CR-7TJtm0
244.
by hedgehog 7mo ago
I'm not sure about the rest but a significant problem with high frequency tool calling (especially in training) is that it breaks batching.
245.
by hedgehog 7mo ago
If you ever did want to wrap this in code processing untrusted images there's a library called "glycin" designed for that purpose (it's used by Loupe, the default Gnome image viewer). https://gnome.pages.gitla
246.
by hedgehog 7mo ago
I keep hoping I'll be able to get my hands on some silphium to see what it's like.
247.
by hedgehog 7mo ago
Or peppers. Hungary without paprika!
248.
by hedgehog 8mo ago
Someone did a screen capture of a SGI "System Tour" app that explains how a lot of the UI worked: https://www.youtube.com/watch?v=sg8fiA9TrRk
249.
by hedgehog 8mo ago
The MS-S1 also has 2x 10Gb Ethernet, the value of the slot is going to a card with more bandwidth + RDMA support.
250.
by hedgehog 8mo ago
I saw that, it's pretty strange.
251.
by hedgehog 8mo ago
Minisforum MS-S1 is the same chip but has a PCIe slot suitable for a network card.
252.
by hedgehog 8mo ago
There are a bunch of tutorials on how to use GRPO to fine tune a small Qwen. Depending what you're doing LoRA or even just prefix tuning can give pretty good results with no special hardware.
253.
by hedgehog 8mo ago
To be fair, in the era when resolutions like 512x342 or 640x480 were common overlapped windows were quite useful.
254.
by hedgehog 8mo ago
To be fair to all involved I'm committing a bit of a category error here, a company of any size is not so much a coherent entity as a flock of individuals with their own motivations. As the popular saying goes: the company ships its or
255.
by hedgehog 8mo ago
Let's assume for the moment the developers are doing about the best they can for the goals they're given. The slowness is a product decision, not an engineering problem. They are swimming in cash and could write a check to solve t
256.
by hedgehog 8mo ago
Claude is slow on a $5k MBP too.
257.
by hedgehog 8mo ago
Not strange, for the kind of applications models at that size are often used for the prefill is the main factor in responsiveness. Large prompt, small completion.
258.
by hedgehog 8mo ago
Mac user since the 90s, 10.14 Mojave seemed like the last release I liked and where the apps I used still worked correctly (or at all). Ironically Apple has already broken most of the apps I cared about, and for the rest LLM tools give enou
259.
by hedgehog 8mo ago
My Thinkpad's fingerprint reader worked out of the box.
260.
by hedgehog 8mo ago
Leopard, which means I've been using it for almost 20 years. The first ten were pretty good.
261.
by hedgehog 8mo ago
I don't know what's happened to Time Machine, with the capabilities available in APFS etc it should be much better if anything. But it's not. Thankfully the failure mode I've experienced didn't lead to data loss but
262.
by hedgehog 8mo ago
It was rumored to be in the original prototypes and cut before launch. I don't know why. They also have very restrictive device limits per account / family and installing apps across accounts is a huge pain. I've mostly given
263.
by hedgehog 8mo ago
There is this: https://code.claude.com/docs/en/devcontainer It does work but I found pretty quickly that I wanted to base my robot sandbox on an image tailored for the project and not the other way around.
264.
by hedgehog 8mo ago
Robots coming for todsacerdoti's job.
265.
by hedgehog 8mo ago
Most of the work I'm aware of starts from the perspective of optimizing inference but the implication that pushing the lessons upstream gets mentioned here and there. Not All Models Suit Expert Offloading: On Local Routing Consistency
266.
by hedgehog 8mo ago
I'm not sure what you're responding to. What I'm describing is my actual experience using Claude, and what I'm hoping for is that they'll spend something like two engineers for a quarter making the app more pleasant
267.
by hedgehog 8mo ago
I don't have links handy but there is active research in this area.
268.
by hedgehog 8mo ago
Oh definitely. The AMD past product just stuck 4x m.2 slots onto the board. Today that approach would be 50-60 GB/s read speed which would be useful enough for something that any of the vendors could build with existing components.
269.
by hedgehog 8mo ago
As a user I would trade fewer features for a UI that doesn't jank and max out the CPU while output is streaming in. I would guess a moderate amount of performance engineering effort could solve the problem without switching stacks or a
270.
by hedgehog 8mo ago
SD is obviously the wrong interface for this but "High Bandwidth Flash" (stacked flash akin to HBM) is in development for exactly this kind of problem. AMD actually made a GPU with onboard flash maybe a decade ago but I think it w

More ›