Rodney D. Gilbert

@Punch_Taylor @sudoingX Thanks. Is wsl better than just using native llama cpp windows cuda binary? Sorry I'm very new to this

4090 datapoint, WSL2 Ubuntu CUDA 13.2, your exact flags + Q4_K_M: ./llama-server -m Qwen3.6-27B-Q4_K_M.gguf -ngl 99 -c 262144 -np 1 -fa on --cache-type-k q4_0 --cache-type-v q4_0 three warm runs on "yo" with thinking auto, system fully idle: - run 1: 42.83 tok/s - run 2: 43.18 tok/s - run 3: 43.33 tok/s - avg ~43.1 tok/s VRAM at 262k provisioned: 23.0GB / 1.1GB free of 24GB. tighter than your 21/3 split — WSL2 + cuda driver reserves eating ~2GB of headroom. native linux would likely give that back. so 4090 + WSL2 = +8.3% over your 3090 native baseline. roughly tracks the bandwidth gap (1008 vs 936 GB/s). bare metal linux on a 4090 should land higher still — would estimate 45-48 tok/s range for someone running native. side observation worth flagging: a single youtube tab in chrome dropped these numbers to ~39.9 tok/s in earlier runs. ~7-8% throughput cost from the browser competing for CPU/scheduling on the WSL side. anyone running this on a daily-driver PC should close everything before measuring.

this was supposed to be a normal evening, then i saw on the timeline that qwen 3.6 27b dense q4 weights from unsloth are live and i could not sit still. compiled llama.cpp with cuda on the single rtx 3090 at 2am from bangkok, launched with the exact same flags that crowned 3.5-27b dense the undisputed king six weeks ago. q4_k_m, 262k context, q4_0 kv cache, flash attention on, single slot, no quant tricks, no dynamic ggufs, no turbo, just the straight cut to get a clean baseline. first pass said "yo" to the model as a warmup. it ran a six step thinking chain to formulate "yo what's up how can i help you today". full reasoning visible in the web ui. thinking mode goes hard, even for a greeting. the number improved. 39.82 tokens per second on the first real generation. march baseline on this exact hardware was 35.3 flat across every context size. that is a 13 percent speed bump. same card, same quant, same every flag, only the model changed. pure model level efficiency on ampere. the model is actually faster at the token level on consumer silicon. 262k context fills 21 gigs of the 24. three gigs headroom for prompt fill. fresh session, zero cache, honest baseline. next i am pushing context, probing the vram ceiling, finding the sweet spot on this card. then autonomous agent tasks on hermes agent using the same prompt that 3.5 dense one-shotted in march. same octopus invaders test, same hermes agent harness, same single 3090 hardware, one model against the ghost of its predecessor. the king might be changing hands.

@RSDCDN Dang. Have you tried to reach out to Leon?

My RGB modded NES died. Sigh. I just ordered the V5 RGBkit for it as I was running the 1.4 for over 10 years. Debating if I even bother modding a new motherboard. Or just sell the parts off. Keeping old hardware going is tiring. My neogeo is dead and I can’t find anyone to fix it

@TakiUdon_ @KarlLetting I ordered the grey one with the dock together, but later ended up splitting them into two per your email (and paying the extra shipping). Should I ask to merge them back together at this point, as it seems they'll line up anyway?

@KarlLetting As long as your order includes a Dock, it will ship together with it.

Merry Christmas. We are still doing QA and shipments today.

@spiderwebsoft Oh shoot 😅

@rodneydgilbert I'm on Playstation. We don't get nice things. :)

I'm about 7 bosses into Silksong, and I think the difficulty debate about it is messed up. People conflate difficulty and just being annoying and eating time. Yeah, it's a hard game. Took me 30 tries to beat my most recent boss. But it's fair and predictable and I learned and won and had fun. The difficulty is generally fine. Tough but fair. The problem is that the game wants to waste my time! Long runbacks. Farming ammo. Farming money. Meeting encounters before I can reasonably beat them. (Designers, please restrain yourself to doing this trick once per game.) To finish the game, I have to commit to wasting, let's say, 20 hours on tedium. And I don't want to do that. So I'll play a while more, and I TOTALLY got my 20 bucks worth. But my playthrough is winding down real fast.

@leonkiriliuk @davepl1968 Apparently it's a watermark for AI.

@davepl1968 What's up with all AI generated images having a yellow undertone? Are we starting to see a model collapse?

Gemini's take on an era correct car advertisement...

@HSVSphere Yeah this is great. I do it with nix-darwin. If only this could be done with full screen transitions.

look mom no animations

@dhh @edimoldovan Embarrassed to say that when I saw "Remember to remove USB installer" I took that as a "you can remove it now before continuing".

@edimoldovan Hot damn!! I think that's a new world record. What were the specs on this install?

@dhh is this a good time?

@HSVSphere @samoletkopetko Hi HSVSphere, is your nixos config available for viewing? Cheers

@samoletkopetko NixOS, niri + a WIP Quickshell config, with all portals etc properly configured. I'd recommend Fedora KDE for someone who wants a system that just works though

Omarchy is ass (it's built on top of arch and doesn't even manage to set up a DE properly) and people treat DHH's uninformed opinions about the Linux Desktop as gospel, but it is nice that people are ditching MacOS, I'll just pray that when they encounter usability issues they

@MisterAddons Super cool. Why is a toggle needed for the last test? Any reason to not have "ppu reset behavior" default ?

The NES core was just updated and now scores 128/128 on Accuracy Coin's latest benchmark. You have to set PPU Reset Behavior to NES in the core's OSD options or else you'll get a 127/128. Open Source FTW. Thanks, Kitrinx!

@thdxr Wish I could disable the full screen transition animation

it is so hard to disable all animations in macos

@RSDCDN Emulation is king 👑

They couldn’t do a fresh render of the pre-rendered videos? Looks outta place now… muddy mess.

I’ve played galaxy numerous times. In 4K on dolphin. On launch day on my Wii. Still own my launch day copy…. Yet still excited for this.

@WadjetEyeGames Use code "Plant Master" and save now!

@GabrielPyron Very nice! What's the display?

Sexy#crt photo is approaching fast on your timeline! #parodius #parodiusDa #konami #240p #15khz #retrogamer #retro #retrogaming #scanlines #pixelart #pixelperfect

@DealsCDN Thanks. Unfortunately not working for me. Discount isn't applying at checkout.

Price applied at checkout.

Great Deal - INIU USB C Cable, 240W [2-Pack 6.6ft]INIU USB C Cable, 240W [2-Pack 6.6ft] Site: Amazon CA 🇨🇦 Price: $8.56 💰 Savings: $9.43 (54%) 💵 amzlink.to/az0GWyw4DqB3y

@HackerVilela I would suggest windows 11 LTSC. No tpm required. MS bloat greatly minimized

My main laptop has a 7th gen Intel processor. 7700HQ @ 3.80 GHz. Great performance. But for business reasons, it cannot run Windows 11. It's sad considering I'm a software developer that does a lot of stuff on retro hardware. I know it can run Windows 11, but they chose to not.

@EliasDaler Super cool! Does this require "tricks" to fool the copy protection checks?

Use this 100% authentic PS1 boot video to do a little trolling when someone says "It's NOT called PSX..."

@dhh Does sleep work on T2? That's been an issue for a while on any Linux distro I believe.

Omarchy 3.0 should install without any additional work on the majority of the pre-M MacBooks. Apple might have given up on them, but many are still fine machines, and Linux can make them run much nicer.

@ES_DE_Frontend Shockingly, Final Fantasy III(6) is also "co-op" ! Well, only in battle, P2 can control certain characters.. but still cool.

Many people love "Secret of Mana" from 1993 but how many people have completed it in Co-op mode? How many didnt know you could play 100% of the game in co-op without P2 dropping out or being sidelined?