- Apple started almost without rapid examples and achieved surprising results
- Starchat-Beta was pushed to an unknown territory without a clear guide
- Almost a million Swiftui programs that work after repeated iterations emerged
Apple researchers recently revealed an experiment in which an AI model was trained to generate user interface code in Swiftui, despite the fact that there were almost no examples of swiftui present in the original data.
The study began with Starchat-Beta, an open source model designed for coding. Its training sources, including the collection and other collections, almost did not contain rapid code.
This absence meant that the model did not have the advantage of the existing examples to guide their answers, which made the results surprising when a stronger system finally emerged.
Creating a personal improvement loop
The equipment solution was to create a feedback cycle. They gave Starchat-Beta a set of interface descriptions and asked him to generate Swiftui programs from those indications.
Each program generated was compiled to ensure that it is really executed. The interfaces that worked were compared to the original descriptions using another model, GPT-4V, which judged if the output coincided with the application.
Only those who passed both stages remained in the data set. This cycle was repeated five times, and with each round, the cleaner data set was feed again to the next model.
At the end of the process, the researchers had almost a million Swiftui samples working and a model called UICODER.
Then, the model was measured both with automated and human evaluation tests, where the results showed that not only worked better than its base model, but also achieved a higher compilation success rate than GPT-4.
One of the surprising aspects of the study is that Swift Code had been excluded almost completely from initial training data.
According to the team, this happened by accident when the subprocess data set was created, leaving only scattered examples found on the web pages.
This supervision rules out the idea that Uicoder simply recycled the code he had already seen; Instead, his improvement came from the iterative cycle of generation, filtering and resentment in his own exits.
Although the results focused on swiftui, the researchers suggested that the approach “would probably generalize other languages and tool kits of UI”.
If so, this could open routes for more models to train in specialized domains where training data are limited.
The perspective raises questions about reliability, sustainability and if synthetic data sets can continue scale without introducing hidden failures.
ICODER was also trained in carefully controlled conditions, and its success in broader environments is not guaranteed.
Through 9to5Mac
