Do we understand how neat LLMs are?

In about an hour on Saturday morning, I had Claude Code:

- Write, train, and test a Bayesian classifier

- Write, train, and test a 2-layer neural network from scratch in Rust

- Write, train, test, then optimize a from-scratch conv-net in Zig

- Run detailed benchmarks on several different scenarios comparing Rust vs Zig performance

Then this morning, I had it create my own Zig-based static site generator in about 20 minutes. Works perfectly.

This tech can be head-banging when writing production code, but it's pretty incredible for exploration.

Do we fully grasp the possibilities?