Akka reports AI porting results across 65 open source projects

Akka says structured specifications changed how often its AI ports landed well on the first try, while model choice shifted speed and token use.

Akka reported a spec-driven AI porting experiment across 65 open source projects that it used to measure how structured specifications and model choice affect delivery quality, cost, and speed. In the first tranche, the work took 99.3 hours and used 9.41 billion tokens, and Akka says 57 of the 65 ports ended with better code size or performance. The company found that more detailed specifications improved first-pass results, while missing context still caused trouble with cross-component decisions, and it plans follow-up work on areas like interface enumeration, test ingestion, provenance tracking, differential testing, and adversarial testing. It also said the lighter model completed ports in 61 minutes on average, compared with 120 minutes for the larger one, even though the larger model used about 40% fewer tokens. Validation combined the original unit and integration suites with checks for serialization, security, error handling, PII, idempotency, and architectural boundaries, and performance varied widely by project, including a reported 143,333x gain for Dify and about 100x slower results for Netflix Metaflow.