Luc Steels: «Science is a cultural activity, it’s like arts: it gives us insights into the world»

La versión traducida al castellano de esta entrevista está disponible en nuestra revista impresa dedicada al pecado #JD13

Defining Luc Steels (Belsele, 1952) simply as a scientist, would be way to poor when we are describing a huge-experienced Belgian linguist who likes art —he wrote an opera and a theater play, besides doing some performances in his youth—, and went to the US to study computer science and artificial intelligence at the MIT with heavyweights like Marvin Minsky and Seymour Papert. Founder and first director of the Sony Computer Science Laboratory, this researcher tries to identify and understand the origins and evolution of language using models and simulations in which several robots develop autonomously their own communication system.

I’ve read that the thing that fascinates you the most is the making of meaning, a topic that has become a thread through all your work over the years. Is it a question that remains unsolved?

Yes, this is my main topic. It’s a topic that traditionally artists and the people of the Humanities are concerned with, but I try to approach it from the viewpoint of building artificial systems for which we could say that they are not only able to treat information, but to create new meaning. I consider the problem of meaning to be the big limitation for today’s Artificial Intelligence. But where are we regarding AI? On the one hand, AI research is clearly very advanced. People don’t know it, but if you use your smartphone or any search engine there’s AI technology behind it. In my opinion the fundamental limitation has to do with meaning. Almost all of the applications that we see today avoid meaning. Let me give you an example: You could now go to Google Translate. Everybody who has used it has had some time an amazing experience because the translation is good, it’s what you expected, but some other times it’s ridiculous. The question is, why is that? Well, it has to do with the way its systems work. What they do is they have access to very big databases of human inputs and they process that information. And I don’t say understand it, I say process it. In the case of translation what they do is they have access to texts where there is a known translation. Then they will pair little bits of texts from the source with a little bit of the text. Then they pick these little bits, which are called engrams, and find them back in your text. Then they take this little bit from the text and they puzzle it together. But they don’t understand the text, they don’t know what it is about. And they don’t try to do very deep linguistic analysis, they have no clue. That’s why I say they don’t use meaning, they purely use the information processing. This is true also for search on the web, because the data bases look for keywords. It’s incredible the scale in which this is happening, it’s just amazing. And we all use it, it’s useful. I use it, and I am happy, but we have to realize about the limitations of it. And so all these people who talk about artificial humanoids taking over, and that we are almost there in ten or twenty years, ignore this fundamental problem that we have. This is a point I wanted to make about the current limits of AI.