0.5 Why Python runs AI, and where it does not belong

Python is slow. A plain Python loop is often 30–100× slower than the same loop in Go or Rust. And yet nearly every model you will use was trained from a Python program. Both statements are true because of one design:

   your Python code            ~1% of the run time:  "what to compute"
   ─────────────────────────────────────────────────────────────────────
   NumPy / PyTorch / JAX       C, C++, CUDA, Triton kernels:  "computing it"
   BLAS, cuDNN, NCCL           ~99% of the run time, on all cores and the GPU

When you write loss = model(x).cross_entropy(y), Python spends a few microseconds deciding which kernels to launch. The GPU then spends milliseconds doing billions of multiply-adds. Python's speed stops mattering when each Python line triggers that much native work. This is why the rules for fast Python are "keep the loops in the library" (vectorize, Ch 31 and 53) and "never loop over individual numbers in Python when a library call can do the whole array".

Python won data science and AI because of:

Readable code close to the maths. y = W @ x + b is the equation.
The REPL and notebooks. Data work is exploratory: load, look, plot, change one thing, look again.
A foreign-function interface that C and C++ authors liked. NumPy (2006) wrapped decades of Fortran and C numerical code; every later library built on NumPy's array.
Network effects. Once scikit-learn, PyTorch, and Hugging Face were Python-first, the papers, tutorials, and hiring followed.

Where Python does not belong, or needs care:

Situation	Problem	What teams do
Low-latency network services at high request rates	interpreter overhead per request, GIL (Ch 27), memory per process	Go or Rust for the hot path; Python behind it or offline
Single-binary CLI tools shipped to many machines	needs an interpreter and dependencies	Go/Rust, or a container, or a tool like `uv tool` / PyInstaller
Model inference in production at scale	Python adds latency and complexity around the kernels	serving engines written in C++/Rust/Go (llama.cpp, vLLM's core, TensorRT-LLM, Triton) with thin Python or none
Mobile and browser	no CPython	export the model (ONNX, GGUF, Core ML) and run it natively
Anything where a type error must never reach production	dynamic typing	strict type checking (Ch 18), tests (Ch 20), or a compiled language

A common split, and the one this Wiki's projects use: Python offline, for data extraction, cleaning, training, evaluation, and export. Go or Rust online, for the services, the agent harness, and serving the exported model. You will see this split in Ch 60–61, and the AI guide's distillation project follows it: train in Python, run somewhere else.