iiinova — where your digital self keeps evolving

bus @bus.bz · 19/09/2026

There is indeed root access on the Frame. Source: youtu.be/_PEq2T_iAs4?...

Terminal in SteamOS on Steam Frame showing sudo commands

1100

bus @bus.bz · 31/07/2026

He's alive! www.da.vidbuchanan.co.uk/blog/slop-cl...

da.vidbuchanan.co.uk

Slop Cleanup Duty | Blog

112

bus @bus.bz · 29/07/2026

I agree with HF's reasoning for file limits too. Seems like a case of "perfect" being the enemy of "works perfectly fine as it is, and has zero issues or needless complexity whatsoever, and everyone understands it".

File size: In the case of uploading large files (e.g. model weights), we strongly recommend splitting them into chunks <200GB each.. There are a few reasons for this:

Uploading and downloading smaller files is much easier both for you and the other users. Connection issues can always happen when streaming data and smaller files avoid resuming from the beginning in case of errors.
Files are served to the users using CloudFront. From our experience, huge files are not cached by this service leading to a slower download speed. In all cases, no single file will exceed 500GB. I.e. 500GB is the hard limit for a single file size.

100

bus @bus.bz · 29/07/2026

Why? Split files have been very useful for me. Like when re-downloading only specific shards of an updated quant. Or when loading a large model when I have more free VRAM than RAM. These could be solved without shards, but they aren't now. Also making HF the next GitHub surely won't backfire.

130

bus @bus.bz · 15/06/2026

To get anywhere close to reasonable speeds you'd need a power-of-2 number number of Sparks for tensor parallelism, and a 200GbE switch (or a PP=2 TP=2 config without one). This review has numbers for TP=4 for the large Qwen and TP=8 for Kimi, not great: www.servethehome.com/big-cluster-...

servethehome.com

BIG AI Cluster Little Power the 8x NVIDIA GB10 Cluster

We built a big AI cluster that used relatively little power. The 8x NVIDIA GB10 cluster shows what scaling this platform can do

120

bus @bus.bz · 10/06/2026

DeepSeek's upcoming vision-capable models will supposedly be good at this, there are examples of them solving mazes and tracing paths: www.k-a.in/Thinking_wit...

Path tracing section of the linked paper, connections being found in the jumble of lines

060

bus @bus.bz · 05/05/2026

So is the workaround to run modprobed-db to find every needed module, load all of them at boot, and then set kernel.modules_disabled=1?

010

bus @bus.bz · 03/05/2026

M2.5 didn't quant particularly well (noticeable benchmark degradation at Q4, so likely much worse at realistic long context work), and M2.7 is even worse (same quanters had trouble even creating their quants). Try the full quality hosted model first if you haven't used it yet.

Messages by mratsim:

"MiniMax M2.7 quanting is problematic, people get high KLD on vllm/sglang NVFP4 or NaN/infinity in llama.cpp lang across many independent quantizers"

"The NaN issues that seem to appear in the llama.cpp (unsloth, AesSSedai and Ubergarm seem to have hit that when doing imatrix) makes me think the model has been overtuned and does not quantize well.

And @luke seemed to be worried about KLD as well in this thread"

010

bus @bus.bz · 30/04/2026

Still not great precedent. I received mine a few days ago and flashed Crosspoint just fine. But I would have had zero interest if I knew there was a chance it'd be locked down.

030

bus @bus.bz · 14/04/2026

I wish someone made a thing like this but with a slightly more capable SoC with networking (maybe just a Pi Zero W) and a slightly expanded keyboard. Would be perfect as a pocket SSH terminal.

110

bus @bus.bz · 12/04/2026

M2.5 is seemingly also still missing weights, but the repo size is the same, as is the number of uploaded shards, so I don't think it's an issue.

124/126 weights uploaded, 230GB repo size

130

bus @bus.bz · 09/02/2026

They literally have (years ago) and that is what makes doing private data correctly so difficult. Multiple people have deployed fully self hosted complete Bluesky stacks, it's not some theoretical architecture

100

bus @bus.bz · 09/01/2026

It is funny, but the missing context is that you wrote all of that yourself and asked to be included. This garbage list generated enough discourse as it is.

PR by iximeow asking to be included in the slopware list

Commit in that PR with the funny text, written by iximeow

1130

bus @bus.bz · 24/06/2025

Can upload it somewhere else if bsky compresses it to death

1170

bus @bus.bz · 03/06/2025

Potentially not their fault: bsky.app/profile/bad-...

1160

bus @bus.bz · 06/02/2025

fellas, is it "heroic" to write desktop software the way it's meant to be written?

160

bus @bus.bz · 12/01/2025

as a lifelong TIS-100 evangelist, i'm happy with any reason that gets people to play it. great game.

020

bus @bus.bz · 20/12/2024

I'm sticking it to the man by hovering over thumbnails and watching tiny videos in their entirety despite this block.

050

bus @bus.bz · 26/11/2024

I don't remember that ever being the default? I registered an alt account days ago and that option is disabled on it. And searching for posts about that option I only see people describing how to enable it.

130

bus @bus.bz · 07/06/2024

The modern Popcorn Time is called Stremio with Torrentio addon. Works great even on Android phones and TVs right from the Play Store.

030

bus @bus.bz · 22/05/2024

Update: it was legit, swapped for $2500.

110

bus @bus.bz · 22/03/2024

My disappointment is immeasurable, and my day is ruined.

5000.0 FLT claimed :fire_emoji:
(with fine print saying you claimed a locked version that's not actually FLT)

130

bus @bus.bz · 20/03/2024

Got nothing to post and nobody to post to, ideal situation really.

110

bus @bus.bz · 06/10/2023

indy 💀 wtf happened?

130

bus @bus.bz · 11/05/2023

I asked that question a few times. Turns out their conviction falls apart real quick once they're cornered into admitting something like "yeah it's a shithole that has no future and you couldn't pay me enough to live there but I agree with what they're doing"

120

bus @bus.bz · 30/04/2023

can't wait to be put on an "evil crypto grifters" blocklist for using the one financial system that didn't kick me out when the walls of dictatorship closed in on me. it's so remarkably privileged to be anti-crypto

1203

bus @bus.bz · 29/04/2023

need to drop a zero or two from the limit to get the real demoscene started

020

bus @bus.bz · 29/04/2023

50KB text posts? in this economy?

120

bus @bus.bz · 18/04/2023

dear god no, lack of a GIF search engine is a massive feature. 90% of their usage can be replaced by reaction emojis which wouldn't take as much space as a paragraph of text. this is why GIF culture is widespread on TG (that had them long before reacts and stickers) but not on Discord

120

bus @bus.bz · 14/04/2023

That makes Mastodon Arch?

020