Weekly AI Digest Issue 11: Recommendation Algorithms, Video Generation Advances, and AGI Research

Weekly AI Digest Issue 11

Published on January 19, 2025, this newsletter presents the latest AI developments.

Market and Voices

1. Xiaohongshu’s Recommendation Algorithm: NoteLLM

With an increasing number of foreign users, Xiaohongshu has become an international social platform. Its new recommendation system, NoteLLM, compresses each note into a token using a specially designed prompt, applies generative‑contrastive learning (GCL) to capture note co‑occurrence, and employs collaborative supervised fine‑tuning (CSFT) to generate timely, relevant tags.

Commentary: NoteLLM delivers more intelligent and personalized UGC recommendations, enhancing user experience as the platform globalizes.

Industry Solutions

1. Chinese Text Generation in Video AI

Recent progress shows Alibaba’s Tongyi Wanxiang surpassing other vendors in generating Chinese subtitles in videos. Comparative tests of three video‑AI systems (Kuaishou’s Kele AI, ByteDance’s Jimo AI, and Alibaba’s Tongyi Wanxiang) demonstrate superior fidelity, motion realism, and camera control in the latter.

The new version also improves complex motion stability, camera movement flexibility, physical simulation, and supports diverse styles and aspect ratios.

Technical details reveal a self‑built variational auto‑encoder (VAE) and a deep image transformation (DiT) architecture, with a spatio‑temporal attention mechanism that enhances modeling of dynamic contexts.

Commentary: Domestic video‑AI models are iterating rapidly, potentially outpacing earlier releases such as OpenAI’s Sora.

2. Keras Founder’s New AGI Initiative

François Chollet and Mike Knoop have launched the Ndea lab, focusing on program synthesis driven by deep learning as a path toward artificial general intelligence (AGI). The approach combines “System 1” deep learning with “System 2” program search, aiming for strong generalization from few examples.

Ndea’s mission is to integrate pattern recognition and logical reasoning into a unified AI system that can accelerate scientific discovery.

The team includes Chollet, a highly cited AI researcher and Keras author, and Knoop, former Zapier co‑founder and AI strategist.

Valuable Technologies

1. Google Veo 2 and Imagen‑3

Google announced Veo 2, a video generation model with higher fidelity, accuracy, and camera‑control capabilities, and Imagen‑3, an advanced text‑to‑image model. Evaluations show Veo 2 produces more realistic motion and better adheres to physical laws compared to competing systems.

Official benchmark results (MovieGenBench, 1003 samples) indicate a strong preference for Veo 2 among human evaluators.

Reference: arXiv:2403.1744