artificial intelligence

In machine learning one of the hottest topics recent has been deep learning. While the computer learns in this method with multiple levels of neural networks it's quite far from how we as humans learn.

There are different levels in our learning as well. Sometimes we just pick up random trivia or light topics. Other times we might need to study a bit more when we encounter a problem we can't yet solve with our current knowledge.  Or we want to do something new and need to start from the basics.

Today, while I was writing a casual AI application I suddenly realized I was using a machine learning model to help me code it. For a while I have been testing out an autocomplete tool called Tabnine which is based on the GPT-2 model trained with millions of source files from Github hosted open source projects.

While the application I was writing won't probably change the world it still made me think. Was I using the AI to write the application, or was the AI using me to write a better version of itself.

Some time ago I read an article about causal AIs. While in the end it was an interesting article I was initially disappointed of it. As I did not read the title correctly the first time. I was in fact looking forward reading about casual AIs instead of causal ones. A little bit of dyslexia there.

So there were no casual AIs to read about. But that got me thinking what would such an AI be like. Is there already such applications in existence?

Imitating human intelligence in software has taken big advancements in the past half century. But despite all the efforts we aren't that much closer to match the human intelligence artificially.

There are two main parts of human intelligence: the conscious reasoning and subconscious sensorimotor skills. The latter usually thought trivial to be replicated due them being rather simple mechanical actions.

Generative adversarial networks are a machine learning method where two neural networks are competing against each other. The generative network tries to create an output based on the initial data and the discriminator network tries to detect if given input is real or generated by it's adversary. Repeating this process while feeding the results back to the network keeps improving both networks and in turn improving the quality of generated content.