Dan Haskin, a DevOps engineer, ran into an expensive problem when he tried building an RSS converter tool in Lisp with AI assistance. Using OpenRouter and the Goose CLI, he burned through $10 to $20 in minutes with barely any progress to show for it. The AI struggled with REPL-based development, spun its wheels constantly, and produced code he ended up rewriting by hand. Then he switched to Python for a companion tool called tmux-repl-mcp, and the difference was stark. Cheap models like DeepSeek and Qwen churned out working code fast. He built the whole thing in a day or two. Same developer, same task complexity, wildly different outcomes. AI models have seen mountains of Python and relatively little Lisp. That training data gap now has a price tag attached. Haskin found himself paying for noise and signal together, and Lisp generated a lot of noise. The REPL workflow that makes Lisp joyful for humans, with its tight feedback loops, actually works against how AI agents operate. These models generate big batches of code in high-latency API calls. They don't iterate the way a person does at a terminal. Add the parenthesis problem, where models routinely fail to balance S-expressions correctly, and you get an experience that feels broken. The community isn't giving up. Bruce Hauman, the creator of Figwheel, is building a Clojure MCP server specifically to handle parenthesis balancing for AI-generated code. The idea is to take the model's raw output and run it through a parser that fixes syntax before it hits the REPL. It decouples the AI's reasoning from its mechanical failures. Other developers report better luck skipping the REPL entirely and feeding models detailed, documentation-rich prompts. Haskin is now considering rewriting his project in Go. Language popularity has always had network effects. Now it affects your AI bill too.