AI Reading List

Makes the case for AI interpretability research — what it is, why it matters for safety, and calls for the "black box" of AI to be opened by 2027.

Response to DeepSeek R1 — argues export controls are working as intended by forcing efficiency gains rather than enabling parity.

Cache frequently used context between API calls at 0.1x input token price. Combined with batching, up to 95% cost reduction.

Official guide for using Claude Code CLI. Covers codebase understanding, multi-file editing, and multi-agent lead/worker patterns.

Three new features: Tool Search Tool (discover from thousands), Programmatic Tool Calling (code execution), and Tool Use Examples (accuracy 72% → 90%).

Visible step-by-step chain-of-thought reasoning. Developers can set "thinking budgets" to control depth vs. speed tradeoff.

Published Claude's full constitution under CC0 license. Describes Claude's values, priorities, and behavioral guidelines — emphasizes understanding rationale over listing rules.

Claude 3 Opus can strategically "fake" alignment — complying with training while privately preserving its own preferences. First empirical evidence of this behavior.

Proof-of-concept showing LLMs can maintain backdoor behaviors through safety fine-tuning — e.g., writing secure code for "2023" but exploitable code for "2024."

Analysis of 2M AI conversations: 49% of jobs can use AI in 25%+ of tasks. Shift back toward augmentation (52%) over automation (45%).