RLHF

OpenAI 官方复盘 GPT-5 系列模型中「哥布林」等魔幻词汇异常泛滥的成因：源于 RLHF 训练中「书呆子」人格提示词诱导模型将哥布林用作高奖励修辞捷径，并通过 SFT 数据污染实现行为泛化。

Rohin Shah argues that while AGI safety risks deserve attention, catastrophic misalignment is not inevitable, and prosaic alignment techniques are likely sufficient to prevent worst-case outcomes, especially since current concerns like deception are not default behaviors in real training.

Cursor leverages sparsity in RL training weights to transmit only deltas, reducing 1TB model sync traffic by 20x for lossless, fast global transfer during active training.

本文对比RLHF与DPO两种方法，揭示大语言模型如何通过偏好学习提升帮助性，解析训练三阶段及技术局限性。

ChatGPT、Gemini和Claude在架构设计上存在显著差异，影响其性能和使用场景。

InstructGPT is a system fine-tuned from GPT-3 that demonstrates how human feedback can transform a capable language model into a far more useful and aligned assistant.

Even assuming AGI requires a new paradigm, applying Lindy's Law suggests it may emerge within 3 to 5 years, so current AI development risks shouldn't be underestimated.

In the AI era, Markdown dominates due to high token efficiency and model preference, but HTML is emerging as the superior output format for interactivity and visual fidelity.

StepFun launches StepAudio 2.5 real-time voice model with paralinguistic perception and personalized interaction capabilities.

Unpacks the pivotal moment when OpenAI's core members were expelled from the precursor to ChatGPT due to a clash with Anthropic's co-founders, outlining the causal links between technical路线 and corporate governance.