FORTH? Really!?
Summary
The post argues that associative and concatenative languages like FORTH may be better suited for transformer architectures than top-down approaches. It introduces the concept of sideways passing joins and a parity-tree benchmark to compare prefix versus postfix notation across models, reporting that postfix and the Opus model perform best when thinking is enabled. The piece blends ideas on language design, token generation, and potential database optimizations through automaton transformations.