Corrigibility cannot be the sole alignment target
Corrigibility — an AI's willingness to be corrected or shut down — cannot be the only alignment target, because realistic training mixes it with other values and leaves gaps that pure versions cannot close. — lesswrong.com
Z-Anime is a full fine-tune of Z-Image
Z-Anime is a fully trained anime-focused model family built on Alibaba's Z-Image, with distilled variants that speed up generation on low-VRAM setups. — huggingface.co
RelayVSR repackages upscaling models for ComfyUI
RelayVSR offers repackaged upscaling model files that drop into ComfyUI's standard folders for local use. — huggingface.co
Qwen3.8 is the Qwen open-model family's most capable model
Qwen3.8 is the Qwen open-model family's most capable model, a Qwen-Max-class release built for coding, professional work, and long-horizon agentic tasks. — huggingface.co
Neuron-based interpretability is more tractable than assumed
Tracing which neurons activate for a concept gives researchers handles for steering and probing language models, making neuron-based interpretability a far more tractable path than previously assumed. — lesswrong.com