Naz Bent

@yacineMTB Qwen 3.5 works fine on my rig🤷‍♂️ x.com/bent302/status…

These AI models are very bad at programming. If you trust them, you are going to have a very bad time

@steipete After going local I rarely even touch chat. x.com/bent302/status…

you said GPT felt like bread, so we stopped serving carbs. 🥖❌🦞

@larsencc Got rid of my chat sub a while ago and anthropic is really testing my patience now too. I think I’ll just stick with Manus and my local AI rig if this keeps up x.com/bent302/status…

Not a good look for Anthropic. Though I like Claude Code, I don’t understand this decision and think it’s really bad.

I’m a lot more interested in business, economics, finance, and psychology/human nature than AI as you might be able to tell. Can’t wait to reveal the true scale of what I’m working on. Stay tuned..

@DJLougen @kaiostephens That’s super cool! I’ve just started teaching myself the ML side of things over the past 3 quarters or so, I look forward to taking a look at that data

@Bent302 @kaiostephens Working on modeling errors currently, I want to bake in an inverse error signals more. Will open source that data once I'm done analyzing it more!

Ran @kaiostephens ‘ Carnice-9b Q6_K against @DJLougen ‘s Harmonic-Hermes-9B Q8_0 — both Hermes-native fine-tunes of Qwen3.5-9B. Same tasks, same stack. This one’s closer. And more interesting. 🧵

@gosrum lol. I’ve been running my heater half as often since i started running this thing x.com/bent302/status…

unslothのgemma-4のggufが更新されたので、再ダウンロードしてベンチマーク回してる 昨日からRTX 5090を動かし続けていて部屋が暑すぎる笑

The Queen requests access to the dryer

@AzFlin Yuh x.com/bent302/status…

niggas'll drop 10 bands on a rig just to run local models for 2 days then switch back to chatGPT

@DJLougen @kaiostephens The results surprised me as well in that regard. Benchmarks like this are what helps this community grow, I cant wait to see both of your V2s if y’all decide to make them. Thank you for your work as well🙌

@Bent302 @kaiostephens Thank you, the delegation stuff is good to know. It seems the model is a little too eager to do stuff on its own. Will need to think more about. @kaiostephens good work homie!

@GOROman I’d love to see what method that is, I think I’ll be rooting a few more over the next year or so

@Bent302 Pixel 8 Pro は root化して使ってたので多分やり方はわかると思う x.com/goroman/status…

Android上にローカルAIを常駐させる仕組みを発明した。いつでもVLM #Gemma4 #AICore #TensorG5

@ProfAdebay I pulled the server out of a scrap bin. The RAM, 3060, and the CPU were purchased Q1 2025. Everything else was this year.

@Bent302 with 3k you built this 2026?

I have no clue how this monstrosity even booted. RTX 3090 FE 24GB 2x RTX 5060TI 16GB (32GB) Quadro RTX 5000 16GB 72GB total VRAM 256GB system RAM 22 core 44 thread intel xeon processor All in cost ~$3000 #opensource #localllama #qwen #ai #nvidia

@kaiostephens It’s a cobbled mess but for someone who got his first computer ever 3 years ago I think I did pretty good 😂

@Bent302 thats beautiful

how much vram do you have?

@GOROman Good luck, I used Magisk to root mine. Worked out pretty well

@Bent302 ありがとう 明日 Pixel 10 Proをroot化してみよう

Harmonic is the better general-purpose agent. Carnice’s edge is specifically orchestration — it delegated correctly and let the 27B do the work. Harmonic tried to do too much itself and ran out of context. They’re complementary. That’s how they’re set up. I will continue using both as separate profiles in my Hermes stack. 🧠 Carnice-9b Q6_K — 35.6 tok/s · Quadro RTX 5000 · 65K ctx ⚡ Harmonic-Hermes-9B Q8_0 — 38.7 tok/s · RTX 5060TI · 32K ctx Full local stack, zero cloud.

Scorecard: •File create + run: Harmonic (cleaner code) •Terminal: Harmonic (better formatting) •Web search: Harmonic (error recovery) •Structured reasoning: Carnice (no wasted tool calls) •Delegation: Carnice (only one that finished) •Direct coding: Harmonic (more production-ready)