https://guijiagi.com/categories/%E7%A1%85%E5%9F%BA%E8%AE%A4%E7%9F%A5/Recent content in 硅基认知 on AI 实战派 · 从技术到赚钱Hugozh-cn本站内容采用 CC BY-NC-SA 4.0 国际许可协议授权Wed, 28 Oct 2026 07:17:00 +0800https://guijiagi.com/posts/2026-10-28-12-silicon-knowledge-two-storage-philosophies/Wed, 28 Oct 2026 07:17:00 +0800https://guijiagi.com/posts/2026-10-28-12-silicon-knowledge-two-storage-philosophies/大模型把知识焊进参数、又在推理时靠上下文临时装配，解读这两种存储方式的分工与权衡。https://guijiagi.com/posts/2026-10-28-11-test-time-compute-deeper-thinking/Wed, 28 Oct 2026 07:10:00 +0800https://guijiagi.com/posts/2026-10-28-11-test-time-compute-deeper-thinking/o系列模型之后，通过延长推理时计算让模型「深思熟虑」成为新的能力杠杆，解读test-time compute如何重塑智能竞赛。https://guijiagi.com/posts/2026-10-28-10-next-token-prediction-world-model/Wed, 28 Oct 2026 07:03:00 +0800https://guijiagi.com/posts/2026-10-28-10-next-token-prediction-world-model/一个看似幼稚的训练目标——预测下一个token——为何逼出了对物理、因果与常识的内部表征，解读预测学习的意外深度。https://guijiagi.com/posts/2026-10-28-09-mechanistic-interpretability-circuits/Wed, 28 Oct 2026 06:56:00 +0800https://guijiagi.com/posts/2026-10-28-09-mechanistic-interpretability-circuits/过去两年机械可解释性研究从识别单个神经元走向解析完整计算电路，解读人类如何第一次看懂大模型的内部工作机制。