Always insightful, always fun
The author reflects on the shifting mental state and market sentiment of 2024, noting the waning hype for AI applications, the importance of research over sheer scale, and the evolving role of multimodal LLMs and scientific scaling laws in shaping the future of AI.
This article introduces LAST, a novel online learning approach that updates recommendation models instantly at serving time, addressing real‑time learning challenges, re‑ranking complexities, and demonstrating superior offline and online performance in industrial e‑commerce scenarios.
The article analyzes why supervised fine‑tuning (SFT) cannot replace reinforcement learning from human feedback (RLHF), highlighting SFT's lack of negative feedback and backward‑looking capability, and explains how RLHF’s reward model addresses these fundamental shortcomings.
The answer explains that joining a pre‑training team accelerates a newcomer’s engineering abilities through hands‑on work with large‑scale data pipelines, distributed training code, and debugging, while SFT teams focus mainly on data labeling, making pre‑training the more effective path for rapid skill growth.
This article presents a comprehensive technical walkthrough on enhancing large language model mathematical reasoning by reviewing model architectures, introducing mixed CoT‑PoT instructions, generating and filtering synthetic data, and applying multi‑stage training optimizations such as RFT, PPO, and DPO, with detailed experimental results and Q&A insights.
OpenAI has launched low‑cost fine‑tuning for GPT‑4o, offering free daily training tokens, a simple dashboard workflow, and early benchmark results that show significant performance gains, while the community debates the merits of fine‑tuning versus prompt‑caching for efficient AI applications.
This article analyzes how Xiaohongshu addresses the decentralization challenge in its recommendation system by strengthening side‑information usage, integrating multimodal signals across the full pipeline, and implementing interest exploration and protection mechanisms, while also outlining future research directions such as generative recommendation and large‑model‑driven user profiling.
This article analyses the shortcomings of current large language model training methods such as SFT, RLHF and DPO, explains why they incur high data and compute costs, and introduces Intuitive Fine‑Tuning (IFT) with temporal residual connections as a cheaper yet effective alternative that better aligns training objectives with real generation tasks.
This article details Alibaba Cloud's end‑to‑end architecture for search and advertising recommendation, covering the data platform, AI services, feature‑store design, training and inference optimizations, and the integration of large language models for new recommendation scenarios.
This article reviews recent research on extending large language model context windows to millions of tokens, covering SAMBA's hybrid architecture, Contextual Position Encoding (CoPE), RoPE base length theory, Retrieval Head analysis, and the memory‑efficient Infini‑Attention mechanism.
