mr. TIM

timkellogg.me

13 October 2026 23:12

ColGrep-Agent-2B a local open agent that you can pop into Claude Code or whatever to get cheap local search ColGrep: multivector code search ColGrep-Agent: navigates ColGrep and reads files to find things quicker & cheaper huggingface.co/blog/lighton...

Rami Krispin

ramikrispin.bsky.social

13 October 2026 23:12

The second tutorial in the MLX series focuses on setting up an MLX server on a Mac and running an LLM locally👇🏼 theaiops.substack.com/p/getting-st... #llm #ai #mlx

Eugene Vinitsky 🍒

eugenevinitsky.bsky.social

13 October 2026 23:12

I have this deep suspicion of RL finetunes of small models that is hard to shake. The models are just so bad, how can you finetune them to do a task reliably

ve, the ghost of info past

vortexegg.com

13 October 2026 10:00

I can tell you from hands on experimenting that you absolutely can make an LLM design software using a fractal module approach with appropriate boundary testing, and you can do human-centered design and evaluation of the functioning of LLM-generated modules in unbounded contexts—it’s a choice not to

crumb

crumb.bsky.social

13 October 2026 09:26

demolishing my intuitions, llama.cpp only gets 8-10t/s for a 4bit 27b while bitsandbytes + transformers gets me 27t/s. what. why