Weekly AI Digest Issue 2: Video Generation, Large Models, AGI, and LoRA Fine‑Tuning

The issue opens with editors Sun Jie and Yu Qi welcoming readers to the second week of the AI roundup, promising practical insights and future discussions aligned with Zhong’an’s business scenarios.

In the “有价值的好技术” section, the article introduces two domestic video generation models: ByteDance’s PixelDance, which uses first‑ and last‑frame guidance to improve temporal consistency, and Vidu 1.5 from Shengshu Technology and Tsinghua, which adds multi‑subject consistency by unifying visual input and output akin to LLMs. Both models are highlighted for their potential in personalized marketing and private‑domain operations.

The “市场与声音” part examines statements from former OpenAI chief scientist Ilya that the era of scaling model size is ending, shifting focus to inference‑time computation and test‑time compute. It also covers Li Kaifu’s analysis of potential U.S. AGI hegemony, suggesting China could compete through ecosystem‑level advantages, lower inference costs, and super‑applications rather than direct AGI competition.

The “前沿论文” segment presents a MIT study contrasting LoRA (low‑rank adaptation) with full fine‑tuning. Findings show LoRA alters the model’s weight singular value decomposition more aggressively, leading to weaker generalization but enabling low‑cost, rapid adaptation. A practical example details LoRA‑based tuning for health‑insurance claim extraction, requiring only ~150 labeled samples and 10‑20 GB GPU memory for about two hours of training.