Roberto Ierusalimschy reveals that Lua's core principle is being a library for embedding in other applications. This "language as a library" approach dictates its core features, including having no global state and enabling cross-language exception handling between Lua and C.

The dream of hardware optimized for functional programming (e.g., dataflow or SK combinator machines) proved to be a mistake. These machines were essentially hardware-based interpreters. The better approach is to build a sophisticated compiler that translates functional code into efficient instructions for general-purpose CPUs.

Lua's speed advantage over languages like Python isn't just from targeted optimization, but also from its small size. This minimalism allows the entire virtual machine to fit within a CPU's cache, leading to significant, natural performance gains that larger runtimes cannot achieve.

Contrary to the belief that abstraction adds overhead, C++ can achieve "negative overhead." High-level constructs give the compiler more information than raw C code, allowing it to perform aggressive optimizations that result in faster executables.

The viability of RISC architecture hinged on compilers becoming sophisticated enough to efficiently manage low-level instructions and register allocation. This software co-evolution was critical to bridging the gap between high-level programming languages and the simpler hardware.

Instead of building a generic graph compiler, Etched focused on hand-optimized kernels. This approach, similar to high-frequency trading firms, provides maximum performance. It's also future-proof, as they design their tools for AI models to use directly, anticipating a time when AI writes its own kernels.

Bjarne Stroustrup advises against being "too clever." Manual optimizations from the 1990s are often "pessimizations" today because they constrain modern compilers, preventing them from applying more sophisticated optimizations tailored to new CPU architectures, caches, and memory access patterns.

Fal maintains a performance edge by building a specialized just-in-time (JIT) compiler for diffusion models. This verticalized approach, inspired by PyTorch 2.0 but more focused, generates more efficient kernels than generalized tools, creating a defensible technical moat.

The line between compiled and interpreted languages is a toolchain choice—you can interpret C or compile Python. The key hallmark of a truly dynamic language is an `eval` function, which requires shipping a compiler within the runtime to generate and execute new code on the fly.