A reaction to a Rails World keynote argues that large language models and ever-lower compilation targets (Rust, C, assembler) are inevitable, but the high-level Rails notation should remain the editable source of truth. A concrete experiment compiled Basecamp’s Campfire down to a single C file: the Rails app is roughly 60K tokens and fits entirely in an agent’s context, while the compiled C is about 4M tokens and ~70× too large. Changing a feature like has_many :boosts requires about 20 coordinated edits across 13 Rails files versus over 1,220 edits in the C, and the compiled output erases the original intent, multiplies opportunities for memory-safety bugs, and forces decompilation rather than maintenance.
A viable source of truth for agents must fit in context, state each decision once, preserve intent, produce deterministic results, and be familiar to models. Treating prompts as the canonical source would need a precise program-like DSL, which is effectively what Rails already is: a compact, conventional specification of web apps. Performance need not be sacrificed by keeping Rails at the top; benchmarks compiling unmodified Rails code show large wins - examples include 85-140× throughput with 96% less memory in one workload and 7-8× throughput with 92% less memory for Campfire - because most runtime cost is repeatedly resolving high-level declarations rather than application logic.
Summary generated by AI from the linked article. hn.today is not affiliated with Hacker News or Y Combinator.