Ai Can Write Its Own Computer Program
For many, it’s the material of nightmares: machines capable of continuously refining themselves. What if they turn malevolent? Will they enslave humanity? Fortunately, given the current status of machine learning research, we will not have to worry about such a scenario for quite some time.
That said, some excellent foundational steps are currently being made. Researchers at MIT have recently developed SketchAdapt, an AI capable of synthesizing short programs. SketchAdapt was trained on over 10,000 program examples and involves a unique two-step generation approach that has not been previously employed.
First, a deep learning neural network uses pattern matching to generate the overall structure of the desired program. The neural network develops the program to a level that it is confident with and leaves placeholders where it is unsure of what to do. Then, a symbolic search algorithm is tasked with filling in the placeholders and developing a program that works correctly. This happens through an iterative trial-and-error approach, repeating until the program works as intended.
SketchAdapt has been used to create programs that can manipulate strings, process lists, and convert math problems into code. The researchers emphasize that SketchAdapt won’t replace programmers but should be able to assist them with writing mundane code. Indeed, rather than replacing programmers, this technology should free them up to work on more complicated projects where man-power is currently lacking.
For more advanced code-generation models, AI ethics will certainly be center-stage. Even now, with programmers paying close attention to data bias, models still often generate biased predictions that can adversely impact minorities. Bias in code-writing AI will be even more damaging, since it can leak into any code generated by the AI. Thus, it will be essential that we build code-generation AIs without any internalized biases.
Though currently very basic, code-writing technology will have far-reaching implications as it matures. One exciting sector where it will undoubtedly play an essential role is healthcare, generating highly efficient algorithms that can detect, treat, and cure disease. While this is clearly a long way off, it will definitely materialize soon enough that it may impact many of our own lives.
Code-writing AI might seem like a step in the wrong direction—a step towards a HAL 9000 robot overseer. However, with proper oversight and ethics considerations, this technology surely has far more potential to help humanity than it does to harm it.
Automation and the Rebirth of American Retail
Why Choose Machine Learning Over A Traditional Financial Advisor?
Interview with Astronaut Scott Kelly
Tesla's Augmented Reality & Virtual Reality Founder Tyler Lindell Opens Up
Supersonic Travel: The Future of Aviation
On Black Holes: Gateway to Another Dimension, or Ghosts of Stars’ Pasts?
Lockheed Martin Confirms the SR-72 – Son of Blackbird Will Reach Anywhere in the World in One Hour
Shedding Light on Dark Matter: Using Machine Learning to Unravel Physics’ Hardest Questions
Interview with the Inventor of Amazon's Alexa
Aquaponics: How Advanced Technology Grows Vegetables In The Desert
The World Cup Does Not Have a Lasting Positive Impact on Hosting Countries
The Implications of Machine Learning on Condensed Matter Physics & Quantum Computing
Written by Daniel DiPietro, Edited by Matthew Durborow & Alexander Fleiss