A widely cited post suggests that dynamic languages are more token-efficient for LLMs because they omit explicit type declarations, making code more compact. For instance, there was a meaningful gap of 2.6x between C and Clojure, while the array language J dominated with an average of only 70 tokens.

However, research indicates that performance on trivial tasks does not generalize to complex problems. While dynamic languages may appear more efficient in simple benchmarks, results become mixed during "ultra effort" tasks. In these complex scenarios, the massive efficiency gains claimed by dense, obscure languages often disappear.

Ultimately, the data suggests a weak positive correlation between language popularity and performance. While obscure, dense languages like J might offer theoretical token advantages, mainstream languages remain a safer bet for most users. This is likely because AI labs focus more synthetic RL data generation on popular languages rather than niche ones, making mainstream languages more reliable for complex coding tasks.