Loading activity…
@zimingliu11
Assistant Professor @Tsinghua_Uni CollegeAI. Founder & Chief Scientist @MetaCircle_AI Science of AI, AI for AI, AI for Science, KAN.
AI
A fixed output shift helped later in training, but @mufan_li disputes the authors’ explanation.
AI
A research paper reports training and validation losses diverging sharply in models with over-encoding n-gram memory.
AI
A researcher calls the vast majority of neurons in large models legible; another says interpretability has far to go.
AI
A paper on video models says a model can generate physically incorrect motion even when the correct motion remains available.
AI
Ziming Liu from Tsinghua University outlines three stages of AI model evolution in a post and blog.
