The project is titled DeepCoder and was built by Microsoft Research’s labs in Cambridge in the UK. As first reported by New Scientist, the AI uses program synthesis to write its own software by joining lines of known code together.
Unlike previous attempts to make a self-writing AI, DeepCoder can create functioning programs in a fraction of a second. Whereas rival systems compute every combination of lines of code to determine the best result, DeepCoder simply looks at a list of inputs and outputs for each line. It then figures out the code required to generate the output value from the given input.
When a successful match is found, the algorithm saves the result for later in case it encounters a similar problem again. It can then immediately recall lines of code it’s previously used to solve problems, speeding up the process next time.
DeepCoder is also capable of learning on its own, effectively taking existing proven lines of code and repurposing them in its own output. Although it doesn’t directly copy current systems – contrary to other media reports – DeepCoder can look through its catalogue of potential code to identify solutions that might work. In the same way a human programmer may keep a notebook of useful code snippets, DeepCoder can peruse its notes and mark potential matches as snippets to investigate further.
The developers behind DeepCoder are keen to expand this in the future, potentially allowing the algorithm to scan repositories of source code created by humans to identify alternative ways of solving a problem. The wealth of code accessible through online libraries such as GitHub could be used by DeepCoder to improve its skills.
DeepCoder remains very limited for now. However, the rise of self-writing computer programs has significant implications for the future of technology. Autonomous vehicles could create new scripts to handle emergencies on-the-fly while error recovery mechanisms could automatically guide malfunctioning programs back towards operation.
In a more meta role, Microsoft also sees self-coding code as a way to lessen the burden on programmers. Developers could focus on the more complex parts of new systems and leave their AI companion to solve the minor tasks.
Significantly, DeepCoder generates human-readable code, unlike previous program-writing AIs. At present, DeepCoder is limited to around five lines of code but the limit should lift in the future. When it does, it could become a new component in development workflows, taking over the tedious tasks coders face each day.
DeepCoder isn’t the first self-coding AI but it’s still a significant step forward in the history of the technology. Last month, AI startup Gamalon detailed its own algorithm, based on a predictive model capable of working out facts related to input data. The purpose is different to DeepCoder though, aimed at expanding data analysis systems rather than actually writing usable code.
