Posts

dlt-proof-writing is an Agent Skill (for Claude Code, or any Anthropic-Agent-Skills-compatible runtime) that automates the bookkeeping side of writing a proof in deep-learning theory. Writing a proof splits roughly in two: finding the structure and picking the assumptions on one side; drawing the dependency graph, looking up citations, getting the format consistent, running lint, doing the read-through-twice-and-fix pass on the other. The skill takes the second half. I keep the first. ...

The setup. Thinking LLMs spend extra compute between <think> and </think> tokens before emitting an answer; pass@1 climbs with thinking length. This was a launch headline when OpenAI o1 and DeepSeek R1 shipped; by the time GPT-5.5, Opus 4.7, Gemini 3.5, Qwen 3.6, and Kimi K2.5 arrived, the curve had become table stakes — every frontier model has it, nobody markets it anymore (Snell et al. 2024; OpenAI 2024; DeepSeek 2025; Muennighoff et al. 2025). The mechanistic question — what’s the rate, and what determines it — is less settled. This post adds one angle. ...