Will AI speak a language of its own?

Worried AI might take your job, or even rule the world? According to Geoffrey Hinton, the Godfather of AI, the reality is scarier than you think. The truth is the way AI has progressed has a lot of parallels with how machine translation evolved. If many once dismissed the idea of AI becoming super-intelligent, so did translators… until statistical machine translation was replaced by neural. Statistical systems relied on patterns in bilingual texts, while neural systems can actually learn meaning from context, just as modern AI does.

Listening to Hinton, I couldn’t help thinking back to my Linguistics lectures at Uni. They taught us Ferdinand de Saussure’s theory of language as a system of signs. Neural networks, however, do not explicitly encode signs. Hinton gives a good example of using a word that doesn’t exist in a sentence, yet we are able to deduce its meaning by context. This is because we associate a word with a big set of active features. Take the word “May”, for example. It could be the name of the month, a surname or even a modal verb. We understand its meaning by the way it interacts with other words in the same utterance.

The alarming part of this podcast is Hinton’s fear that AI may develop its own language for thinking, a language no humans would understand. This certainly can’t be ruled out considering the frightening rate AI is developing at now, faster than anyone ever thought. So the challenge isn’t whether it will be useful, it clearly will be, but whether we can keep control.

The way AI operates is it uses data fed into it and learns from that data, including harmful data. So it’s vital that we model good behaviour, similar to what we do when raising a child. Sadly, this isn’t what we are doing at present. During the Cold War big nations managed to collaborate to avoid nuclear catastrophe. Today no such safeguards exist for AI. At the same time it is getting increasingly persuasive in changing people’s opinions and influencing our decision-making.

The question is no longer whether AI will change our world, but whether we can shape it before it shapes us. A big challenge for everyone involved in machine learning now.

Will AI rule the world?

[Tip: If you have used AI for translation of your content and wish us to review it, please email us enquiry@talkrussian.com]