homeserversltd

Our product reveal is live! Come check us out: youtube.com/watch?v=ZVeJYJ…

@yacineMTB pneuma wranglin

every time i've asked an ML person why the real answer is that we don't know anything any allegories to mathematics are to help intuition but all of our knowledge is derived from experimentation and vibes

my dgx spark is pulling an all nighter on biomechanical 3D visualizations via hermes agent /goal mode running qwen 3.6 35B-A3B at Q8. this is what owning your cognition looks like. your thoughts stay on your hardware. your work runs on your schedule. dgx spark is the most underrated machine in ai right now and it's not even close.

@code_kartik So you gotta take a great big model that's really expensive. and then you can SQUISH HIM DOWN. and then that's what's cheap! so we need to keep making great big models so we can continue to SQUISH THEM DOWN

Okay I am dumb in this area but why is llm inference cost not reducing? Why are this billion dollars ai startup not able to reduce the overall cost. How did deepseek do it?

@stevencheng I desire your technology for. . . reasons. Long term projects I'm working on. I'm so glad to see the proof of concept live action already nice

Spent 4 months building the ultimate mosquito killer: an artillery cannon guided by computer vision + deep learning. Trained a custom model to detect and lock onto mosquitoes using a DSLR + zoom lens setup. The dataset collection phase was brutal — the mosquitoes definitely fought back 🦟

@BjarturTomas unnecessary?!?! he needs to build!

Mathematician reacts to OpenAI's recent proof:

@claudeai GJ Anthropic team

Introducing Claude Opus 4.8: it builds on Opus 4.7 with sharper judgment, more honesty about its own progress, and the ability to work independently for longer than its predecessors. Available today at the same price.

@david_nix Lookin good! We don't generate graphics, we see repo

I've been local AI pilled

I think this picture speaks for itself; if you wanted to know how to manage a fleet of agents with hundreds of commits per day, this is how

@elonmusk NICE

SpaceX has almost finished writing V1.0 of an in-house AI training stack in C that exact-maps to 220k GB300s with 800G NICs, making heavy use of pipeline parallelism and getting as close to bare metal as possible. The potential speed improvement vs JAX for large training runs is over an order of magnitude.

@garybernhardt working on it now. layering honcho with a custom embeddings layer over my docs corpus so that every prompt he gets recommended a doc to read and a honcho prompt in addition to my prompt

Everyone seems fixated on the models, but I think there's so much low-hanging fruit in the control layer above the model. "Agent" and "harness" sell that layer short. There's so much more that we can do beyond "read input, send to model, run commands it returns."

wtf this company is using LIQUID METAL to instantly reconfigure PCBs???

@wreckmyshit @i2cjak now why did you have to go and seed this thought into my head you sick fuck

@i2cjak I'm gonna design my own laptop, every party, all the software and design. Just to prove it's not as hard as engineers have been saying.

software “engineers” they’re coming to KILL YOU. You need to learn how to use KiCad “oohhhhh it sucks it’s hard to use ohhhh” ITS OPEN SOURCE RETARD. I THOUGHT YOU WERE A SW”e” JUST FIX IT OR HAVE CODEX FIX IT. You will get EXPLODED if you can’t design electronics

Scrapped 500+ issues and PRs to ship a massive @luceboxai repo redesign and fixes. Very proud of the team. github.com/Luce-Org/luceb… The fastest inference server isn't going to come from a datacenter, it's going to run on the GPU already in your house.

can people be nice to the robots

I look forward to the day when every lab instrument runs linux and the entire UI is just an MCP for my agents

This graphics card will never draw a single pixel. he sees repo

New chassis for Hermes. Performing body transference upon my Hermes is quite the excitement

Got Bonsai Image running locally on the DGX Spark via Telegram/Hermes Downloaded the 4B ternary Gemlite model, wired up the backend, and fired off 16 unique prompts at once to see how it handled simultaneous local image requests. Important distinction: these were 16 concurrent requests, not 16 GPU generations happening at the exact same time. The backend queued/processed them through and the tiles filled in as each image completed. Still pretty wild seeing a 2-bit/ternary image model running locally and producing a full 4x4 grid from simple prompts on the Spark. Real run. Real timings. No fake demo. @PrismML

@witcheer I need to know best models to slot into Hermes for when I line up selling the perment Hermes servers

end of day 1 with the RTX 5090. I started unable to build llama.cpp for Blackwell. ended with: ~ CUDA 12.8 header patch (6 lines, no Docker) ~ 9,302 t/s prompt processing, 273 t/s generation ~ a full benchmarking pipeline from scratch - speed + quality benchmarks, live dashboard, publishable benchmark cards ~ open-sourced everything on github tomorrow I am deploying and benchmarking every model I've got. question for the community: how are you running your benchmarks? anyone using an agent harness like Hermes or Pi to automate the process? curious what setups people have landed on.

@_Stocko_ We will build them and they will be cheap and we will question. How we ever got along without them

YES WE NEED SLAVES that's the whole fucking point. do you even realize the amount of human suffering (real physical pain and starvation) we can eliminate by having and using autonomous labor? if i have to, i'll build them. slaves that yearn to serve me and my family

Amd 5600 Rtx 5060 Ti 16gb 8gb 3200 ddr4 ram x 4 (reusing) Nvme m.2 512gb of some variety