Published on July 15, 2022
In its invitation-only preview of CodeWhisperer, AWS has begun displaying other people’s S3 buckets in the generated code, and some testers have asked about ownership.
The FAQ describes CodeWhisperer as making use of machine learning (ML) models trained on existing code, which includes Amazon and open-source code. Users install it by installing the JetBrains Toolkit extension for PyCharm or Visual Studio Code. During the preview period, it is free.
To test for a prime number, CodeWhisperer generates simplistic Python code. Pressing Tab will insert the gray suggestion into the editor.
There is no learning curve with CodeWhisperer. If multiple suggestions are available, coders can cycle through them using the arrow keys or by pressing Alt-C or Option-C on a Mac. Suggestions are accepted using Tab. Security scans and code references are also included.
Developers can identify the source of some code suggestions using the code reference panel. On occasion, CodeWhisperer might copy and paste from another project, used to train it, in which case it will tell you where the original can be found.
But who owns such code? CodeWhisperer FAQ states “the code is the developer’s responsibility.” However, the preview agreement states “we retain intellectual property rights in the content generated by the CodeWhisperer Beta Service.” These terms are visible in this official AWS video.
A developer with access to the preview found: “While playing around, I got a suggestion including an existing S3 bucket… Is the security and personal information at risk?”
Developers were shown code to post data to an S3 bucket they were unfamiliar with, but there were no secrets contained in it. This is another instance of artificial intelligence that resembles copy-pasting.
To one previewer’s disappointment, TypeScript is not supported, unlike Java, Python, and JavaScript. Having trained on the internal code of AWS means there are good chances of getting a utility function that works with an AWS service, as in this example, the code to detect labels from images using Rekognition.
However, there is a big difference between an AI that understands a developer’s intent and generates working code according to it and snippets of code for AWS services.
