Conventional programming code is highly specific and fixed for a particular task. In machine learning, we have rather simple and universal neuronal network layers, which obtain their specific capabilities from learning data.

Reinforcement learning and Recurrent neural networks have a self-modifying capacity: the self-modification of internal neuronal network data in the hidden layers in order to change behavior (data is the new code). AutoML is even modifying its Neuronal Network Architecture.

But there is also already code that writes the next generation of code, and that’s Darwinian evolution: AI will replace programmers by 2040. Machine learning and natural language processing technologies will be so advanced that they will be capable of writing better software code faster than the best human coders.

Google’s AI can create better machine-learning code than the researchers who made it.