Demis Hassabis, British artificial intelligence researcher, cofounder of Google’s DeepMind AI project, envisions a future where robots can actually advance medicine by formulating and testing scientific theories, and where AI can solve humanity’s biggest problems. Does he live in a dream world? Current progress in AI says no.

Current AI technologies are specialized in a very few, usually one field. For example there is DeepMind’s AlphaGo software which have beaten world champion Lee Sedol in a five-game Go championship; this meant a new milestone in computing history. This is the software that learnt using 1980’s console games by itself, and ended up playing them better than a human could. The ominous Go game surprised professional Go players and scientists all over the world, because computers haven’ yet been able to do so complex calculations. To put this into perspective, IBM’s DeepBlue software defeated world chess champion Garry Kasparov in 1997, but compared to chess, where a player has an average of 35 possible moves every turn, a Go game provides about 250 possible moves each turn. As Hassabis puts it, “chess is a calculation game, but Go is too complex, so players are using their intuition”.

DeepMind is currently planned to be involved in multiple usage areas, for example medical examinations to help spot signs of kidney problems that are commonly missed, and cause a large number of avoidable deaths. DeepMind could also be useful to improve Google’s virtual assistant system with better recommendations. The latter has a huge impact on Google’s revenues because of their advertising system’s better targeting.

Ethical and philosophical questions

While advancing in artificial intelligence studies is a huge step toward a better future, it also raises the question of safety, but as Hassabis says there is an internal board of philosophers, lawyers and businesspeople who look after the project and that DeepMind “is nowhere near anything we would be worried about”, but if everything goes as planned, the ethics board will have some real work to do in the near future.

Robots blending into society

Meanwhile at Stanford University researchers are developing a computer-vision algorithm that can predict the movement of people in crowded places, therefore analysing and respecting human social conventions. The ultimate goal of the research is for robots to blend into human society and roam around “unnoticed”, without bumping into people every two meters. Jodi Forlizzi of Carnegie Mellon University says “a lot of robots will be working with people and close to people, so we need to understand how they should behave” in such environments.

AI that builds AI

Only a couple of hundred people have the knowledge to push artificial intelligence research forward, and a large proportion of their worktime is spent with trial and error of different theories and practices. Some tech giants are already trying to develop AI that can test algorithms on behalf of the engineers. At Facebook they are building AI algorithms that can help build AI algorithms, their goal is to create new AI models using as little human work as possible. Apart from better ad targeting, face recognition and news feed algorithms, Facebook representatives say this technology could ultimately help engineers to generate audio captions for photos so that the blind can understand what’s in them.

But no need to worry, we are not anywhere near to a frightening story as Skynet’s.

*Picture courtesy of Universal Studios.