DHH, creator of Ruby on Rails, tasked an AI agent with rewriting his Campfire Once application from Ruby on Rails into Rust, Elixir and Go without reading the generated code himself. The experiment aimed to test whether extensive prior experience could compensate for not inspecting AI output, but the resulting rewrites revealed immediate non-functional differences and performance flaws under load.
In the Rust version, stress testing showed a mere 1% success rate for delivering new-post notifications under heavy load due to a tokio::sync::broadcast channel limited to 256 messages. Increasing the channel capacity to 16,384 boosted delivery to about 90%, but the worst-case event latency rose from 11 seconds to over 130 seconds, illustrating a trade-off where higher throughput came at the cost of delayed failure detection and poor user experience.
The Elixir version delivered nearly 100% of notifications at around 1,700 requests per second, but achieved this at the expense of substantially higher memory usage—peaking near 1.8 GB—and experienced HTTP POST timeouts on roughly 23% of requests. These outcomes demonstrate that no language or AI-generated solution eliminates the need to reason about failure modes, resource bounds, and user-visible latency, reinforcing the author’s conclusion that reliable systems still demand thoughtful design decisions.