Toy Models of Superposition：神經網路的特徵壓縮術
打開 IDE。你觀察隱藏層的單一神經元。它同時響應了 ‘Python’ 和 ‘錯誤處理’。這不是 bug，這是特徵壓縮。Anthropic 的 Toy Models 研究揭示特徵稀疏性如何驅動「疊加（Superposition）」現象，解釋模型如何在容量限制下儲存更多概念。
Tim
Beyond Chinchilla: How Do You Balance Training and Inference Costs?
When inference demand is extremely high, following the Chinchilla rule can actually be more expensive. This article explains how a “small model, more data” strategy can find a new balance point in the total cost of training and inference.
Constitutional AI: How AI Feedback Enables Automated Safety Governance
Annotator A gives it a 5, annotator B gives it a 1. The same answer. The bottleneck in traditional RLHF is not cost, but the signal-to-noise ratio. This article breaks down how Constitutional AI uses AI Feedback to build automated safety boundaries, and the cost structure teams need to evaluate when putting it into practice.
Kaplan 2020 論文：Scaling Laws 的演進與情境邊界
2020 年 OpenAI 提出模型大小與資料量的冪次律，建議『大模型、少資料』。兩年後 Chinchilla 修正 compute-optimal 配置，讓資料量的重要性重新被看見。拆解 Scaling Laws 演進，以及它如何影響你現在的訓練決策與成本結構。
Kaplan 2020: Scaling Laws and Their Limits
In 2020, OpenAI proposed a power-law relationship between model size and data volume, which led many teams toward a simple idea: bigger models, less data. Two years later, Chinchilla revised that picture with a compute-optimal view, showing that data volume mattered far more than many people had assumed. This matters if you’re deciding whether the next gain should come from more GPUs, a larger parameter count, or a larger corpus. The history of scaling laws is not just academic. It changes how you think about training efficiency, model selection, and where your cost structure really sits.
Sparrow Paper: Training Trustworthy Dialogue Agents with Rule-Conditional Reward Models
The CI/CD pipeline is red again. You switch back to the logs and find the model confidently inventing an API endpoint that does not exist. DeepMind’s Sparrow paper proposes an RCRM architecture that uses rule-conditional reward models and evidence chains to let the model automatically verify sources while generating. This is not about making the model smarter. It is about teaching it how to prove that its answer is correct.
AGI 的維度：如何透過能力與自主度評估 AI 風險
當 AI 標籤被行銷用語模糊化時，工程師該如何評估風險？本文透過 Performance 與 Autonomy 的正交維度，提供一套實用的部署風險矩陣。
AGI Dimensions: How to Assess AI Risk Through Capability and Autonomy
When AI labels are blurred by marketing language, how should engineers assess risk? This article uses the orthogonal dimensions of Performance and Autonomy to provide a practical deployment risk matrix.
RLHF 的代價：InstructGPT 如何靠對齊而非規模，讓 1.3B 打敗 175B 的 GPT-3
175B 的 GPT-3 輸給了 1.3B 的 InstructGPT。這不是參數的勝利，是『聽話』的勝利。但聽話的代價，是你在部分任務上可能付出能力取捨。本文拆解 RLHF 機制與對齊稅，說明如何在可控性與通用性之間做出取捨。
The Cost of RLHF: Why 1.3B Beat 175B
GPT-3 at 175B lost to InstructGPT at 1.3B. This was not a win for parameter count. It was a win for instruction following. But that gain comes with a tradeoff: when you push a model to stay closer to human preference, you may also narrow its performance on some public NLP tasks. This article breaks down how RLHF works through SFT, reward models, and PPO, then looks at the alignment tax and what it means when you have to choose between controllability and broader capability.