Local-AI on Phil Wheat

Joy Caption Alpha Two

Tue, 13 Oct 2026 09:05:00 -0500

JoyCaption is an image-captioning vision-language model built from the ground up to be free, open, and uncensored, aimed at generating captions for training diffusion models — fancyfeast’s claim is near-GPT-4o captioning quality while staying unrestricted. Alpha Two is the Llama-3.1-based iteration (LLaVA-style), with prompt-able caption formats (descriptive, Midjourney-style, art critic, product listing, tags) and 4-bit quantizations for low-VRAM cards. The linked HF Space is the hosted demo; ComfyUI custom nodes exist for local use.

Yan Desbiens

Tue, 13 Oct 2026 09:01:00 -0500

Blogger — some interesting ideas. AEON is what caught my eye on this one.

Yan Desbiens runs a one-person, one-RTX-4090 local-first AI lab and documents it all. The AEON project: an open-source, deterministic procedural world simulation steered by a local LLM “world-spirit”, with embodied citizens, a teacher-to-student liquid-neural-network mind, and a Three.js dashboard. Also on the site: NeuroArch (background cognition cycles between prompts), ForgeLM (train 30M–500M models on one 4090), UFM (fractal memory so a single GPU runs models that shouldn’t fit), agentos, and genetic architecture search. Unfashionable from-first-principles local AI, written up candidly.

Strix Halo × Qwen3.8 (pwilkin)

Tue, 13 Oct 2026 08:57:00 -0500

A full, reproducible stack for running Qwen3.8-27B on Strix Halo (gfx1151): calibrated IQ4_XS weights, an IQ4_XS DFlash2 drafter, the UMA scheduler ring, and retained-PM4 dispatch on a custom-built ROCm userspace — installed from source into your home directory via one installer command, no system ROCm replacement. Their matched-protocol reproduction claims fastest-on-Radeon-8060S by 13–25% over the Vulkan-derived stacks (~26 t/s decode with ~61% draft acceptance, ~257 t/s prefill), with machine-readable benchmark evidence and a broader community leaderboard table. Directly relevant to the Halo box.

Train Your Own Decision Model with Unsloth

Tue, 13 Oct 2026 08:43:00 -0500

Using Unsloth desktop to train a classifier model.

Unsloth’s guide to turning ordinary LLMs (Qwen, Gemma, Llama) into decision models like their Jev/Laya/Clef: the fine-tuned model reads input, scores the options you provide, and returns a choice with a calibrated probability instead of generating text. LoRA (rank 16, 4-bit, lr 2e-4, 1–2 epochs) lifts accuracy from roughly chance (30–37%) to 73–81% on decision benchmarks (typed-decisions, BANKING77, CLINC150), with test sets decontaminated against training. Runs through the Unsloth Desktop app’s Train tab (macOS/Windows/Linux) or code via FastDecisionModel, then serves the model under their Decision API. Modest VRAM: a 0.8B model trains in 4 GB in ~40 minutes.