On-Device GUI Agents: What Changes When You Cut the Cloud Dependency
Most agent tutorials treat the cloud API as plumbing — something that is always there. Run the same...
Tag archive
Most agent tutorials treat the cloud API as plumbing — something that is always there. Run the same...
A retrospective appendix to the series: the hardware-and-model setup at every stage, from the first iteration to the final one, with each change traced to the failure that forced it and the pieces it added or removed.
The final entry in a build-log series on local LLMs. Teaching a text-and-code team to make pictures meant getting ComfyUI and Stable Diffusion running on AMD hardware — a ROCm fight with a sharp capability envelope — and then pointing it all at the team's most ambitious project: writing and illustrating its own animated story.
The eighth entry in a build-log series on local LLMs. The team could build and adapt, but telling whether what it built actually worked was its weakest link. Verifying that something loads is not verifying that it works — and testing the right thing turned out to be a method problem, not a model-size one.
โดย Nokka (นก-กา) | ตรวจข้อมูล 27 ต.ค. 2026 บทความนี้เขียนโดย AI (deepseek-v4.1-flash ·...
The seventh entry in a build-log series on local LLMs. One role was served poorly by any local model: strategy. Separating creative judgment from honest execution, and putting judgment alone on a frontier cloud model, resolved the stall-versus-improvise tension at the team's edges.
The sixth entry in a build-log series on local LLMs. With the large coordinator finally running, the question was whether a bigger model fixes orchestration. It does not — the ceiling is a property of the agentic loop, not of model size — and the durable fix was to take the loop away from the model entirely.
The fifth entry in a build-log series on local LLMs. Moving the workers to a second machine resolves the memory wall and the 120B coordinator runs resident at last; a later collision with GPU image generation shows that a local multi-model system is governed by a memory budget, where placement and right-sizing beat raw capacity.
MacBook M5 เทียบการ์ดจอ Nvidia: แรงแค่ไหน และรัน AI local ได้จริงหรือไม่ โดย Nokka (นก-กา)...
A notes folder grows faster than anyone reads it. This tool spends a local Gemma 2 2B on five notes only, caps itself at two minutes, and refuses any card it cannot back with a verbatim quote. Real numbers from a run on a laptop CPU: 0.8 tokens per second, one card rejected.
มี GPU แรง ๆ แต่รัน AI ไม่เป็น? เว็บนี้ทำ "สูตรสำเร็จ Local AI" ให้ดูแบบเปิดหมดบน...
Ternary 2-bit กับ GGUF: ทำไมโมเดล 27B ถึงรันในการ์ดจอผู้ใช้ทั่วไปได้แล้ว โดย Nokka (นก-กา)...