Perfection Not Required? Human-AI Partnerships in Code Translation
summary
Unsupervised NMT model TransCoder translates Java/C++/Python but only passes unit tests about 75% of the time, so outputs often need repair; 11 software engineers evaluate three UI features and conclude imperfect code is still useful. Confidence highlighting steers attention to likely errors and serves as a review checklist, while alternate translations expose plausible variants and help catch bugs. Engineers treat model output like work from a junior developer, prefer rigorous testing, and want feedback loops to retrain and improve the model.
AI-written summary · about
related talks
34mThe Human-in-the-Loop is Tired
01h 30mEffective AI pair-programming
01h 28mFuture of Software Programming with AI Bots at Hand
26mAl and Software: The Good, The Bad, The Scary
26mBuild AI Systems for Discernment, Not Approval
41mAI or Not AI? We Don't (Yet) Believe the AI That Replaces Devs!
46mCode Quality in the AI era, needed more than ever
19mWith great power comes great responsibility in code quality