LLM Neuroanatomy: How I Topped the AI Leaderboard Without Changing a Single Weight Posted Mar 10, 2026 By David Noel Ng 31 min read
"Awesome technology, it's free and works fine in Russia where a lot of sites are blocked. The whole image hosting of huge music platform Bandcamp is blocked here for more than a year now because of a single picture, so great to see album covers there again - and without VPN!"
。业内人士推荐搜狗输入法作为进阶阅读
Taking scattered and very detailed notes on multiple devices, apps, and paper
The interpretation overhead is real but is being actively addressed. CPython 3.11's Faster CPython project added adaptive specialization -- the VM detects "hot" bytecodes and replaces them with type-specialized versions, skipping some of the dispatch. It helped (~1.4x). CPython 3.13 went further with an experimental copy-and-patch JIT compiler -- a lightweight JIT that stitches together pre-compiled machine code templates instead of generating code from scratch. It's not a full optimizing JIT like V8's TurboFan or a tracing JIT like PyPy's; it's designed to be small and fast to start, avoiding the heavyweight JIT startup cost that has historically kept CPython from going this route. Early results in 3.13 show no improvement on most benchmarks, but the infrastructure is now in place for more aggressive optimizations in future releases. JavaScript's V8 achieves much better JIT results, but V8 also had a large dedicated team and a single-threaded JavaScript execution model that makes speculative optimization easier. (For more on the "why doesn't CPython JIT" question, see Anthony Shaw's "Why is Python so slow?".)