Author of the BaiMian large model, offering technology and industry insights.
By replacing the traditional Pass@1 reward with a Pass@k formulation and a lightweight advantage computation, a 7B language model can dramatically improve its performance on math reasoning benchmarks, surpassing GPT‑4o while adding only a few lines of code and minimal training overhead.
This article systematically adapts classic deep reinforcement‑learning techniques—such as multi‑step returns, TD(λ)/GAE, V‑trace corrections, uncertainty‑aware weighting, safety constraints, distribution‑robust optimization, and value‑guided decoding—to improve large language model training and inference, providing concrete formulas, implementation tips, and empirical results.
The article analyzes the poor generalization of supervised fine‑tuning (SFT) for large language models, reveals its gradient as a high‑variance inverse‑probability policy gradient, proposes a one‑line Dynamic Fine‑Tuning correction, and shows substantial gains on challenging math and offline RL benchmarks.
In a WAIC 2025 interview, Zhu Yibo, co‑founder of Jiejie Xingchen, shares deep insights on AI infrastructure, covering its evolution, the need for tight model‑hardware co‑design, cost‑efficiency metrics, industry challenges, and future directions for large‑scale AI systems.
This article analyzes the surprising design choices of the rumored GPT‑OSS 120B model, explaining the rationale behind a 64‑dimensional attention head, the equal hidden and intermediate sizes, and other quirky parameters such as MLP bias and KV‑sink SWA, backed by theoretical formulas and empirical benchmarks.
This article breaks down the hardware bottlenecks of large‑scale LLM training, explains the Roofline performance model, arithmetic intensity, and how computation and communication costs interact on GPUs and TPUs, offering concrete formulas and examples for efficient scaling.
The article introduces Qwen3‑Coder‑30B‑A3B‑Instruct (aka Qwen3‑Coder‑Flash), detailing its architecture, 256K‑to‑1M token context, agentic coding capabilities, installation steps with Transformers, sample code for tool use, optimal sampling parameters, and deployment tips across various runtimes.
The Qwen3‑30B‑A3B‑Instruct‑2507 model, an updated non‑thinking version of Qwen3‑30B‑A3B, delivers significant gains in instruction following, reasoning, multilingual knowledge coverage, and 256K context length, and its performance is benchmarked against leading LLMs across a wide range of tasks.
The article analyzes AWS Bedrock AgentCore and related hardware and software requirements for Agentic AI, covering runtime isolation with microVMs, memory architectures, identity and gateway design, zero‑trust networking, and the challenges of multi‑tenant KVCache and context engineering.
This article compiles more than thirty interview‑style questions and detailed answers covering large‑model fundamentals such as encoder‑decoder trade‑offs, self‑attention versus RNN, context length, tokenization, embedding strategies, FlashAttention, RoPE, prompt design, retrieval‑augmented generation, safety measures, fine‑tuning, and model distillation, providing a comprehensive technical reference for practitioners.
