Facebook AI now fixes bugs like spellcheck corrects typos
Writing got a lot easier when spellcheckers and grammar checkers arrived on the scene. Now Facebook is using artificial intelligence to give programmers a similar boost with a tool called SapFix that’ll debug their software.
‘SapFix can automatically generate fixes for specific bugs, and then propose them to engineers for approval and deployment to production,’ Facebook engineers Yue Jia, Ke Mao and Mark Harman said in a blog post Thursday.
Automation technology in the past has made humans obsolete for such physical tasks as weaving fabric or installing car windshields. With the brainier work AI can tackle so far, though, it seems to be augmenting human abilities rather than replacing humans altogether for chores like translating text, analyzing medical scans or programming.
Certainly Facebook doesn’t see SapFix as replacing human programmers. It’ll free them from ‘the drudgery of tedious bug fixing,’ and perhaps even could encourage more people to take up programming since the AI has their back, the company said.
Not just for catching crashes
And Facebook has hopes that SapFix can handle more types of bugs.
‘In the future, we will be able to automatically fix not just crashes, but all sorts of other problems, such as performance issues …, reduced battery life, high bandwidth consumption and memory-hungry apps,’ the company said.
SapFix works in conjunction with an earlier tool called Sapienz that Facebook developed to automate software testing, a process that can catch software problems early in its development. (Yes, in case you wondered, the SapFix name is derived from Sapienz.)
Facebook plans to release both as open-source software, a move that could help other programmers benefit from the tools.
The company detailed the projects at its @Scale conference for engineers responsible for running the gargantuan computing systems at places like Facebook, Google, Amazon and Microsoft.
Partners for Facebook’s Glow
Also at the conference, Facebook announced progress with a project called Glow designed to make AI software run faster. Chipmakers including Intel, Qualcomm and Marvell are now allies, Facebook engineers Vijay Rao and Nadev Rotem said in a blog post.
AI, which uses machine learning technology loosely based on human brains, today runs on hardware to accelerate both the early training phase and the later ‘inference’ stage where AI actually does useful work like recognizing faces. Glow is designed to speed up both phases.
Glow’s job is to optimize AI systems built with tools like Facebook’s PyTorch. Some of that can be done with general improvements, but some is tuned specifically to the hardware the AI system is running on, so partnerships with chipmakers are important.