Always in the middle of something.
Chasing ideas across ML, AI, and data. Building tools when the rabbit hole gets interesting enough.
Editor's picks
How tool call arguments stream in the Claude API
A look at the input_json_delta events that carry a tool call's arguments in a streamed Claude API response, and what changes when a tool streams them eagerly.
Claude API 串流時的工具呼叫與不完整的 JSON
用 Claude API 開串流的時候，參數會是一段段字串。這篇簡單的看這些片段長什麼樣子，以及程式該怎麼接手處理。
Where the cache_control breakpoint goes in Claude prompt caching
A walk through how the Claude API writes and looks up prompt cache entries, using the docs' example of a breakpoint that never gets a cache hit, and how to check the result in the usage fields.
Claude API 的 prompt caching：cache_control 放在哪裡
來看看 Claude API 的 prompt caching 是怎麼找快取的，以及 cache_control 放在哪裡會影響有沒有正確讀到。
開源專案怎麼看 AI 寫的程式碼：QEMU、Linux 核心與 Debian 的規定
QEMU、Linux 核心和 Debian 都有在文件寫怎麼處理 AI 的貢獻。這篇簡單的看三家各自的規定以及理由。
GPT-6 Astra 評測與價格：跟 GPT-5.6 Sol、Fable 5.1 的比較
我們來看看新模型 GPT-6 Astra 跟上一代 Sol、Anthropic 的 Fable 5.1 在分數和花費上差在哪裡。
The Assisted-by trailer in GCC, the Linux kernel and QEMU
How three projects currently handle the Assisted-by: commit trailer on AI-assisted patches: GCC's policy, the Linux kernel's documented format, and the checkpatch rule QEMU merged in September 2026, with the source text each one comes from.
What GPT-6 Astra Costs Per Token and Per Task
GPT-6 Astra lists at $10/$50 per million tokens against GPT-5.6 Sol's $4/$20. A walk through Artificial Analysis's per-task cost figures for both, and the one benchmark where Astra's lower turn count sits beside a lower score.
Python 3.14's free-threaded build
A look at Python 3.14's free-threaded build through the CPython 3.14 docs and source: installing and identifying it, what happens when it imports a C extension, and what it costs in speed and memory.
bfloat16 vs float16: how two 16-bit formats split their bits
A look at bfloat16 and float16 through their bit layouts, loss scaling, and a small example of low-precision addition.