Writing Lisp Is AI Resistant and I'm Sad

11 hours ago

The author, a DevOps engineer using AI tools like OpenRouter and Goose CLI, attempted to write an RSS reader format converter in Lisp but faced challenges with AI-assisted REPL development.
AI models (Claude, DeepSeek, Qwen) struggled with Lisp in the REPL, leading to high costs and low productivity, unlike with Python where AI performed dramatically better.
To improve REPL interaction, the author created tmux-repl-mcp in Python, simplifying command execution and output retrieval, which integrated easily with existing AI tooling.
Despite the tool, AI still had difficulty with Lisp, resulting in poor signal-to-noise ratio and high costs, prompting consideration of rewriting the project in Go for better AI support.
Issues include AI's reliance on common tools like QuickLisp over alternatives, and the inherent mismatch between high-latency AI APIs and interactive REPL development.
The author reflects that AI favors languages with abundant training data (e.g., Python, Go), making them cheaper and easier for AI-assisted coding, while Lisp remains resistant.
A historical analogy is drawn to Plank Road in Naperville, IL, comparing Lisp's nicety but niche appeal to the efficiency of railroads (like AI with popular languages), illustrating 'Worse is Better'.
The author wonders how Lisp will adapt to the AI era, hoping for future improvements to make AI more effective with Lisp, despite preferring its elegance and fun.

Hasty Briefsbeta