WOW I just updated Halogen and is it even faster now? I'm typing and it's coming back almost instantly!! This is crazy awesome for the Strix Halos Halogen 0.10.2 —> 0.16.2 (GHCR latest, Oct 2 build)

EasyAgent is coming together nicely Still working on Tool Calling but once it's in a place where it's working well I'll release the repo link I was really inspired by @coldniko and what he was able to do with Strata So here I am building an Agent framework Don't worry the Model is still training also

Building a new Agent Framework!! After Using Grok @bot I've realized how bloated and not easy to use Hermes is even though I'm using it So I'm building EasyAgent the Easiest Fastest agent around You can use less than 64k context to run more bots never have to compress because it offloads chat to text to keep going and never stop Still early in Development I'll provide a repo link soon

Grok @bot Pro Tip: Something I learned talking to @SpaceX team Shout out to them for listening and wanting feedback! Don't use Grok @bot to build code point it to a cursor agent and you'll get better cheaper code

This is honestly impressive @coldniko Strata + Qwen3.8-Flash-Next (IQ2_XS) on a Tesla V100 32GB — experimental sm_70 build. ~40 tok/s wall-clock from another box on the LAN (128 tokens / 3.2s). Engine-side decode was ~58–74 tok/s with ~95% expert-cache hit. This is running on an old Z440 with a Single 32GB V100 and only 40GB of RAM I guess it's time to get 128GB of Ram for the Z440s

Question for Everyone When you complete something wtih AI Do you say "I Did it" Or do you say "We Did it"

AI Agent Pro Tip: When your agent is stuck and not making progress, stop the local loop. Have it kick off a grid search that maps the decision space: variables, ranges, probabilities, and expected outcomes. Blind iteration is how agents burn tokens without learning. A structured map of the possibilities is often what unsticks them.

American 🇺🇸 Open Source Lab Day — Log 2 Day 2 of training the 300M ternary model at 100 TPP. Current step loss: 1.71 That's not bad at this stage. We're early in the run (~228K steps), and the curve is behaving. For those asking what the number means: loss = how wrong the model's next-token bet was. Lower = more accurate, less perplexity. 1.71 means it's putting real probability on the right token — learning, not memorizing, not flailing.

WOW The Strix Halo is still smelling like a rose compared to anything else for Price/Performance Glad I have two of them!!

Training is live again!!! Training the 300 Million Parameter with 100 tokens/param for 30B total tokens Heavy coding focus Fully Open Source once this model is finished training we'll release everything we did to make it Hopefully it helps speed up what other people are working on American 🇺🇸 Open Source For The Win Follow along with this link as it trains ultront1live-kqxip0.stacksho…

Geez this almost makes Ternary not worth the time but if I can make a large enough Dense Ternary model it will likely outperform most other things Imagine a 1T Dense model you could run at home and how intelligent it would be We have all seen how good 27B is

American Lab Daily Log — Oct 1, 2026 WE ARE BACK TRAINING. Picked a cheaper on-demand box — 2× RTX 3090 — so the run takes a little longer, but the burn is easier to absorb day to day. This is the grind lane for Ultron T1: from-scratch, vibe-ready, fully open. This will be the first model we release on Hugging Face, along with everything used to create it — pack recipe, code, checkpoints, board — fully open source. Live stats (now) Board: ultront1live-kqxip0.stacksho… Training on Vast.ai Hardware: 2× RTX 3090 · 512 GB disk Cost: ~$0.42/hr (~$10.1/day) Model: Ultron T1 ~300M @ 100 TPP target (~30B tokens) Resume from: ckpt_12000 (~3.15B tokens already seen) Remaining: ~26.85B tokens · est. ~26 days at ~12k tok/s Pack: pack_100tpp_30B upload ~49.5% (29.5G / 60.0G @ ~9 MB/s)

Real Question When Cloud AI Fails are we going to be able to pick up some H200's for like $2k? If hundreds of thousdands come on the market or is China just going to snap them up? I would love to have an H200 cluster at home I would be willing to install a bigger circuit

I just wanted to say thank you my 138 Followers If I haven't followed you back and you want me to let me know If you are looking for verified Followers I'll trade follow for follow to help people reach creator status I'm trying to do some cool stuff with AI I'm in a state of transition right now but should be back working on models soon

Looks like I’m 2 of 10 Harder Faster!!!

When we all buying Huawei GPUs?? Was it a mistake to not sell Nvidia to China? Now they’ve just taken all the designs they stole and made their own chips

This is Breaking news..............the UK has AI Research?

The videos just keep getting better and better

This guy might have something here, haven't tried it yet but hopefully today if not definitely this weekend lots of new inference frameworks coming out with huge improvements over llama.cpp