Intel Corp., MIT and the Georgia Institute of Technology today detailed an artificial intelligence code analysis system that is being hailed as a “solid step” toward teaching computers how to program.

Intel is also looking to turn the system into a coding assistance tool for developers. 

Last year, the chipmaker set up a new group within its research organization to  pursue a concept it terms machine programming. Justin Gottschlich, the group’s director, defines the vision behind machine programming as enabling “everyone to create software by expressing their intention in whatever fashion that’s best for them.” Users would simply have to tell an AI what application they want and the model would do the coding. 

But there are still big obstacles standing in the way of Intel’s machine programming vision. One foundational challenge is that for an AI to write code based on a user’s description, say a plain-English list of desired features, it would first need to understand what it is exactly the user is asking for. In other words, the AI would have to figure out their intent. That’s the area that the system announced today aims to advance. 

The system, dubbed MISIM, can analyze the intent behind two different snippets of code to figure out if they perform similar tasks. This capability represents an early but potentially important step toward enabling computers to interpret user intent in an automated software development context.  

MISIM can, for example, compare two snippets of code written by two different developers to perform the same calculation. MISIM is capable of determining that those two code snippets have the same intent behind them, that is to say they perform the same task, even if the two developers used different algorithms and data types to implement the calculation.  

What makes the system stand out is its accuracy. The Intel, MIT and Georgia Tech researchers behind the project wrote in a pre-print paper that MISIM, with its ability to extract intent from code, is up to 40.6 times more accurate than previous code comparison systems. The researchers tested the system by having it assess 45,780 different programs. 

Machine programming is an “audacious goal, and while there’s much more work to be done, MISIM is a solid step toward it,” Intel’s Gottschlich said.

The system’s accuracy edge comes from an innovation called context-aware semantic structure. According to Intel, that’s what enables MISIM to look at a piece of code and capture what the code does rather than how it does it. MISIM’s context-aware semantic structure creates a high-level description of the intent behind a code snippet that can be understood by a neural network.

Intel is already exploring ways to put the technology to use. On top of boasting impressive accuracy, MISIM is capable of evaluating code while a developer is writing it, a feature Intel wants to harness to give engineers real-time programming recommendations. The chipmaker hopes to use the system for tasks such as finding bugs and performance issues.

A hypothetical MISIM-inspired bug hunting tool would work by comparing a code snippet to other snippets with a similar purpose in order to find differences. If the developer’s code diverges significantly at a certain line, that could indicate a flaw. The same approach could theoretically lend itself to finding performance issues. 

MISIM is still being improved, but Intel has moved it from the research phase to the demonstration phase to explore potential applications.  The chipmaker’s machine programming research group is now in talks with internal software teams about integrating the AI into their workflows. “I imagine most developers would happily let the machine find and fix bugs for them, if it could – I know I would,” Gottschlich said. 

Photo: Unsplash