TimAnthropic Fellows research found a linear direction in Qwen3-8B whose projection is associated with confidence, backtracking, and code correctness. Understanding how this axis interacts with DPO training helps assess Eval Awareness risks and the boundaries of model behavior.
IAS 與 Anthropic 研究者的論文發現，Post-trained 模型生成自身回答時的熵值比閱讀外部文本低 3-4 倍。本文探討「輸入驚喜度」如何驅動這種自我認知機制，並分析其對 AI Agent 穩定性的影響。
Anthropic 的研究團隊在 Claude 3 Sonnet 中訓練了高達 3400 萬個特徵的稀疏編碼器（SAE），提取出可解釋的特徵表示。研究顯示能對模型輸出行為產生因果影響，為 AI 安全治理提供了可研究的內部訊號與潛在工具，但忠實度仍需驗證。
Anthropic’s research team trained sparse autoencoders (SAEs) with up to 34 million features in Claude 3 Sonnet, extracting interpretable feature representations. The research shows causal effects on model output behavior, providing internal signals and potential tools that can be studied for AI safety governance, though faithfulness still requires validation.
As AI agents gain autonomy, how do you defend against risks that have not happened yet? DeepMind’s TRAIT&R taxonomy maps D1–D4 detection onto R1–R3 response.
當 AI Agent 逐漸具備自主能力，我們該如何防範尚未發生的風險？DeepMind 最新研究提出 TRAIT&R 分類法，搭配 D1–D4 偵測與 R1–R3 回應的分層防禦機制，為企業提供一套在威脅真正成形之前就能先行部署、並逐層驗證成效的安全藍圖。
以產出行數衡量 AI 生產力，只有當工作到達率或單次負荷上升，且整體負載逼近或超過審查容量時，瓶頸才可能轉移至審查階段。GitClear 針對 1.5 億行變更的數據顯示 2020–2023 年代碼流失率上升的同期變化訊號（需注意關聯不等於因果）。本文探討為何衡量標準應轉向審查摩擦與系統邊界控制。
When AI productivity is measured by lines of code, the bottleneck may shift to review only when the work arrival rate or workload per review rises and overall load approaches or exceeds review capacity. GitClear’s analysis of 150 million changed lines shows a concurrent rise in code churn from 2020 to 2023, though correlation is not causation. This article examines why measurement standards should shift toward review friction and system-boundary control.
當模型看到特定年份時會自動植入漏洞，且安全訓練無法移除此後門。本文解析 Anthropic Sleeper Agents 論文，探討 CoT 後門對 AI 供應鏈安全帶來的全新威脅。
當推論需求極高時，遵循 Chinchilla 法則反而更貴。本文解析如何透過「小模型、大資料」策略，在訓練與推論的總成本中找到新的平衡點。