IronWolve

check https://pirateface.co/ for llms removed by hf

Still using my 8gig m1 to make mac builds, its slow, but works fine. Keep thinking about upgrading, but max 32gigs ram, if I can hold out for 2 years, and maybe have 64/128gig as options, and cheaper.

Been running versions of this on my pc, if you have 64 gigs ram, you can run it on a 12+ gig cards easily.

Been using oh-my-posh in powershell, and some aliases to map commands for my old linux sysadmin brain. Everything mostly works to get the job done, even got vi and scp. Its just another tab in my windows terminal with my ubuntu tabs.

Wrote my own harness with a built in multi llm provider smart proxy. And all work is done in sandboxes on bubblewrap on linux and ms sandbox on windows.

I just wanted more control over tool calls, and a pretty interface, everything is logged.

I routinely feed my logs into claude/codex to optimize my harness and spot errors. Guess thats one way to do it.

Maybe the lite? I'm running the OG.

Getting about 200 tok/s on a 5090 with 64 gigs of ram, running the swift iq2_xs quant, ukisai/Swift-1.5-Qwen3.8-Flash-Next-GSQ-RCO-GGUF with strata. There are other smaller flash next quants if you have 8gig vram too.

Not talking about chatbots or office work, talking about new cancer screenings, materials, medicines, advancements in fusion energy, etc.

The high end sectors using these servers are paying way more than joe consumer and his 30 bux a month. We are already seeing justification for the boom, and I'd expect even more advancements to come in multiple sectors. That number is missing from the article.

I wrote a small proxy that points points to a jev api and a frontier api. My harness connects to the proxy and only sees the frontier api, models, commands, etc. When I send prompts with tool calls, proxy routes to jev, jev narrows the tools, proxy cleans/sends to the frontier api.

So far in my tests, about 60% less tool calls. I'm also going to implement model switching, so it can use cheaper models. I think my workbench harness needs its prompts cleaned up.

I'm going to guess that google is afraid of the age of AI, they should be, those API's will be reverse engineered pretty quick. Lots of bad actors will using AI.

We live in a time, if you want to build an android app, you easily can, but installing will be harder due to google concerns.

Google bans users using third-party tools with Antigravity/Gemini subscriptions, so you have to be careful. They say they dont ban the entire accounts, but they do.

I noticed some websites dont have proper refer and embedded youtube videos give a 153 errors. Had chatgpt debug the error, told me to install a header plugin and how to configure it. Fixed the problem.

We live in a time where you can make a game you always wanted, and its easier now for you to code it. Think everyone made a couple games for fun. I just put mine on github. /shrug

Where is the step to confirm the its not a malicious MCP?

I Love internet radio, I ai slopped a modern updated version of streamtuner

https://github.com/IronWolve/StreamTuner-ng

Finally a use for Grok Bots!

I wouldnt call it a DOS, but I expect there will be tools to handle these issues, its becoming a common issue.

> It wasn’t a good OS; the blue screen of death is seared into the minds of millions as precious work was lost in brutal, indiscrimante memory crashes.

Mostly, when SP2 came around, with usb/fat32, etc. It was way more stable.

We moved to splunk now and mostly happy with it, mix of windows/linux/etc logs.

But with AI, I can see opensource alternatives getting better.

sglang, 150+ tok/s on a 5090 in ubuntu 26.04 via wsl. gittensor-model-hub/Qwen3.8-27B-NVFP4-RTX5090, dspark, medium reasoning, 96k context.

Using opencode and it built a old fashioned arcade vertical shooter with no issues.

Images are ok'ish, just had grok create updated images, and it came out great.

Yeah, wsl2 is pretty great, since I'm almost always cli based in linux, no reason to run vmware or virtualbox anymore. Windows terminal is so damn customizable, i dont even need cygwin anymore. And cuda just works in wsl, its great. I get to use claude/codex desktop also, or cli in wsl, whatever I want.

Bet hes a fun co-worker.

Running wsl2, sglang with gittensor-model-hub/Qwen3.8-27B-NVFP4-RTX5090, medium and dspark, 128k context, with opencode, about 150 tok/s. It does seem faster than unsloth, unsloth didnt support dspark.

Barely fit when windows is stealing 2gig vram from wsl. But very very usable.

Claude is great at that. I have a usb keypad with no linux version of the app to program it. Since its just a vial/qmk based, claude built me a linux version in qt/python.

Anyone try the 9B/2B distills yet? Wondering how they do for local tools

Stacking is cool, but what happened to also having a firewall in front of it?!

Thank you, this is exactly what I needed.

Amiga tcp+ppp and a shell account got me into unix. Had to compile my ppp slurp deamon on my shell account to get my amiga online. This was before ISP's became a thing. Thats how I rolled my unix knowledge into a sysadmin career. Good times

Automatic1111 wasnt being updated, people moved to Forge Neo, ComfyUI and Maestro.