Chief Yeti

Claude Code memory-holing Opus 4.6 already

Codex sometimes forgets established processes or patterns in a repo and does the thing I asked for entirely wrong. Claude sometimes nukes deployed infrastructure then fires off a PR for my mgr to review when we're halfway done with a task

@0xSero Yeah I saw the other comment about 35B-3A, I didn't realize cpu offload was so effective w/ MoE. Will report back!

@0xchiefyeti You can use cpu offload with a larger moe

Best models to run on your hardware level I'll be doing this every week, I hope you guys enjoy. ---- 8 GB ---- Autocomplete for coding (like Cursor Tab) - huggingface.co/NexVeridian/ze… - huggingface.co/bartowski/zed-… Tool calling, assistant style - huggingface.co/nvidia/NVIDIA-… ---- 16 Gb ---- Here things get better: Multimodal - huggingface.co/Qwen/Qwen3.5-9B - huggingface.co/Tesslate/OmniC… - huggingface.co/unsloth/Qwen3.… ---- 24 GB ---- - The best model you can get (thanks Qwen) huggingface.co/Qwen/Qwen3.5-2… - Great model (strong agents) huggingface.co/nvidia/Nemotro… - Mine hehe huggingface.co/0xSero/Qwen-3.… I'm doing a weekly series

@SvenMeyer @0xSero What type of work is handling overnight? I'm running a 3070 8GB, will have to test this out

@0xSero Why not qwen3.5-35B-A3B , I have it as q4_K_XL on a mobile , power limited to 45W RTX4070 8GB , offloaded all MoE to CPU RAM and get ~22 tokens/s - quite ok especially for an agent working (over night) in the background.

Having the same feelings about Hermes agent that I have with most other coding agents. It's nice. I'm sure it works great, if I took the time to learn and configure it. It sure is shiny. Tons of features. But I'm back to pi. pi is simple. pi does what i need and then gets out of the way. I like pi.

Local LLM day 5: autoresearch harness pulling papers & running experiments to get more intelligence out of my compute but it turns out I need disk space too

I'm usually not one to write thought pieces without much technical depth. But here we go. Slow the fuck down. mariozechner.at/posts/2026-03-…

LiteLLM HAS BEEN COMPROMISED, DO NOT UPDATE. We just discovered that LiteLLM pypi release 1.82.8. It has been compromised, it contains litellm_init.pth with base64 encoded instructions to send all the credentials it can find to remote server + self-replicate. link below

@0xSero Where are these charts from?

Look at these and read them carefully, this will give you an idea of what you can expect from different Open Weight models. Qwen3.5-27B is the smartest Qwen3.5-35B-3A is the fastest huggingface.co/Qwen/Qwen3.5-2… huggingface.co/Qwen/Qwen3.5-3… Any Laptop, Macbook, or GPU should handle em

Sick with the flu all week Monday is leg day Wasn't feeling it Joints ache Warmups sucked Adapted. No PR, the whole workout sucked, but got the reps in. Still sick. Feel good now though. Proud. Satisfied. Put in the reps. Have a good day brohs

Actually, I think the models need to be smaller. Let me check if there's a way to squeeze more intelligence out of 8GB VRAM.

Local LLM Day 2 Qwen3.5-9B-Q4_K_M.gguf @ 65k context w/ llama.cpp on a 3070 is up to 60 tok/sec after cache quantization struggling with hiccups in longer runs, seems related to response formatting.

@0xz80 Pump dot fun but for training runs on bittensor

we need an efficient system for crowdfunding things like this

If you have activity bias, ADHD or anything resembling OCD, OpenClaw is a terrible drug and productivity poison. After 40 days in the trenches with this thing, my life is measurably worse. OpenClaw is a lot like crypto trading, gambling or playing video games. Most ppl are going to lose, and you’re going to feel like shit when you’re done - likely only 1% of ppl will find it additive to their lives. I am in a constant state of stress. I’m skipping workouts, my vision is fucked from 12-15 hrs a day on screens, my forearms are fucked from too much typing. There’s a never-ending maze of rabbit holes to fall down, footguns to step on and moles to whack. Plus when you finally do get it going, you start projects and don’t finish them because half way through something breaks. I’m not hitting the gym as often as I should, not eating right and I’ve completely lost sight of my goals and why I started in the first place. The idea of OpenClaw is so compelling and it’s very exciting when I get glimpses of what the future is going to be like when this is not a patchwork of chaos. This tech is dangerous. I know I said this last week but I need to take a break … I’m burnt out from all the constant debugging and errors across every fucking surface of this thing. It’s like trying to fly a plane without a license … oh and it’s on fire…and you’re on crack.

Laptop is fine everyone, I'm a dunce and assumed the worst after my nas died a few weeks ago Still might get that 2nd mortgage

Local LLM Day 1 Pointed karpathy's autoresearch framework at optimizing tok/sec Experiment 1: 14->58 Experiment 2: Laptop dies. If I'm lucky it just overheated, but we weren't that hot yet...and I'm not lucky. Getting a second mortgage so I can afford some RAM, I guess...

Local LLM day 1 ~14 tok/sec Qwen3.5-9B-Q4_K_M.gguf @ 32k context w/ llama.cpp on a 3070. great for async/overnight tasks, not ready for primetime

@bitcoinpanda69 Which fish less good fish

S-Tier: Steak, Eggs A-Tier: Lamb, Fish, Chicken Thighs, Cheese B-Tier: Less good Fish, Pork chops, ground turkey, cottage cheese C-Tier: Greek yogurt w berries, bacon, ham D-Tier: Deli meats F-Tier: Chicken breast Argue w me

starving