Skip to content

Highlights

Personal stories I keep coming back to.

Latest posts

Notes from the workshop floor.
AI & Agents8 min read

Krea 2 Turbo: Why 16-Channel Latents and ConvRot Beat 24GB GPUs

How Krea 2 Turbo delivers photorealistic chiaroscuro and candid depth on a 12GB GPU using 16-channel Wan VAE latents and ConvRot INT8 quantization.

#comfyui#diffusion-models#vram-optimization
AI & Agents10 min read

Hindsight: Beyond RAG, How AI Agents Build Long-Term Memory That Learns

What is Hindsight? Deep dive into vectorize-io/hindsight: the open-source biomimetic agent memory engine replacing naive RAG with Retain, Recall, and Reflect.

#ai-agents#mcp#rag
AI & Agents9 min read

BFS Head V1.1 Explained: In-Context Head Swap without Plastic Skin in Qwen Image 2.1

Master in-context head swapping in Qwen Image 2.1 DiT with BFS Head V1.1. Eliminate neck seams and plastic skin via surgical 16-layer MLP weight ablation.

#qwen#generative-ai#machine-learning
AI & Agents15 min read

Lanshu AI Presenter Video Explained: Audio-First Architecture vs Viral Hype

Deconstructing Lanshu, the viral AI presenter video skill: audio-first timeline clocks, motion plate separation, billable circuit breakers, and hype vs reality.

#ai-video#ai-agents#codex
AI & Agents13 min read

Pi Agent in Practice: Setup, Custom Extensions & 5 Best Practices

Master Pi Agent from scratch: complete CLI setup, multi-provider LLM routing, custom TypeScript extensions, and 5 battle-tested best practices.

#pi-mono#ai-agents#cli
Updated
AI & Agents11 min read

Qwen-Image-2.1 Viggle Turbo LoRA: 4-Step DMD2 Speed Guide for ComfyUI

Qwen-Image-2.1-viggle-turbo LoRA: 4-step DMD2 distillation cuts inference 10x on consumer GPUs. ComfyUI GGUF setup, LoRA adapter config, and benchmark results.

#qwen#generative-ai#machine-learning
AI & Agents11 min read

Qwen-Image-2.1 ncnn Vulkan: Run 7B DiT on 2GB VRAM Without CUDA

Run Qwen-Image-2.1 locally on Intel, AMD, and Mac GPUs with 2GB VRAM. Master nihui's portable C++ ncnn Vulkan engine without Python, PyTorch, or CUDA lock-in.

#qwen#generative-ai#vulkan
AI & Agents11 min read

OpenCreator Explained: Open-Source AI Video Subtitling & Dubbing

Deep dive into OpenCreator (formerly KrillinAI), an open-source Go & Electron engine that transcribes, translates, dubs, and reframes videos into vertical shorts.

#ai-agents#video-automation#tts
AI & Agents12 min read

Jev Ultrafast Explained: Sub-20ms Decision Engine for Browser AI Agents

Mổ xẻ jev-ultrafast: sub-20ms browser agent speed, atomic DOM extraction, and the brutal truth behind TypeSafe vendor lock-in and cherry-picked benchmarks.

#ai-agents#browser-automation#web-scraping

From the HoangYell galaxy

Sister projects I quietly maintain.