Fast and Hard Code

Reflections on the evolving landscape of software development Original Author: Armin Ronacher | Date: August 22, 2026

There is a recurring joke on Twitter claiming that "programming is solved now." While the absolute truth of that statement is debatable, one reality is undeniable: the hurdle of mastering a specific language's syntax is no longer a primary barrier. The friction that once slowed down human developers is largely irrelevant to AI agents.

Consequently, the specific choice of a programming language has become far less critical than it once was.

The New Paradigm of Language Selection

Because LLMs can translate logic across different syntaxes, you are no longer tethered to the languages you know. If you dislike a particular implementation, you can simply instruct an agent to rewrite it in a different language—even one you have never studied.

This shift means that language adoption is now driven more by marketing and perceived benefits than by developer familiarity.

"As a long-term Rust programmer, it's been fascinating to see a surge of people shipping Rust code who, in the past, would have been deterred by the learning curve."

The "Vibe Shift" Toward Performance

This trend is fueled by two primary cultural shifts:

1. A renewed, widespread desire for high-performance, fast software.
2. The realization that LLMs are incredibly proficient at optimizing code without breaking existing functionality.

These individuals have always been obsessed with speed, and they've embraced the idea of agents writing code. With tools like autoresearch, you don't need to memorize every obscure optimization trick; you can simply deploy an agent to find them (though having the knowledge still provides a significant advantage).

The Rise of "Hard Languages"

We are seeing a growing number of projects that prioritize being small and fast, leading them to choose "hard languages." Interestingly, this includes Zig, even though some of its core community members remain skeptical of AI.

Real-world examples include:

• Cloudflare's Artifacts: Utilizes a Git-protocol engine written in pure Zig, which compiles down to a WebAssembly module of only $\approx 100\text{ KB}$.
• Vercel's fx: A coding agent built with Zig, specifically marketed for its speed and minimal footprint.

It appears that most of these modern, high-performance projects are being developed with the assistance of LLMs.

Breaking the Gatekeepers

The impact extends beyond just the choice of language; developers are now tackling "harder" technical domains that were previously considered off-limits or too intimidating.

Current trends in LLM-assisted low-level work:

• Analyzing DWARF files
• Implementing eBPF
• Writing custom network drivers
• Developing custom cryptography
• Programming for legacy computing hardware

In the past, these fields were often gatekept by a small circle of experts. Now, the barrier to entry has collapsed.

Conclusion: The Trade-off

We can model the current state of software production as a balance between quantity and quality:

$\text{Software Output} = \text{AI Slop} + \text{High-Performance Engineering}$

While we may see an increase in "slop" (low-quality AI-generated code), we are also seeing a new generation of developers who are empowered to demand that their software be fast and compact.

---

Tags: #ai #programming #thoughts Licensed under CC BY-NC 4.0