October | 2017
Originally published in LinkedIn
If you are into castles and dragons, blood and betrayal and men fighting with swords, you are perhaps already addicted to George RR Martin’s television series, Game of Thrones (GoT). I am intrigued by the series for an entirely different reason: Zack Thoutt, a software engineer, fed all 5,000-odd pages of the previous five GoT books to an artificial intelligence (AI) system called a recurrent neural network and out came the next few chapters of Martin’s epic[i]. The language isn’t great but the style is close to Martin’s, say reports, and fans think the scenarios conjured by the AI are plausible. You can take a look at some of these chapters on GitHub and decide for yourself. But the question is: How long before machines can begin to write exactly like Martin, manipulating a cast of over 2,000 across complex plots? Or for that matter, how long before machines can begin to do everything that we ask of them?
Many suggest that Singularity, the point at which the distinction between humans and machines will be erased, could be just around the corner, say by 2030 or 2050. This is because not only are compute capabilities expanding exponentially, but costs are coming down. What could take years and decades in a computer lab – say, the ability to translate a language with 80% accuracy – has become an inexpensive and everyday commercial possibility. Today’s publicly accessible systems can translate languages with almost 90% accuracy. But, navigating the 90% threshold is still a good few decades away.
Robots are getting smarter for sure. Lego makes millions of bricks every hour and packages them with absolute precision using robotics[ii]; robo-assisted surgery is helping make complex procedures safe and cheap[iii]; beautiful homes are being built (literally) overnight by a bunch of automated arms driven by CAD drawings[iv]. These industries – from manufacturing to medicine – are at the cutting edge of technology, busy re-writing business models and human life cycles. But are we close to being run over by machines? Not at all.
First, the ability to do everything requires more than brute compute capability. It requires the exchange of sophisticated and fine-grained knowledge between machines. But machines don’t have a uniform way of collaborating between them. For machines to create a complete sum of all parts is decades away.
The truth is that we don’t even know all the critical factors and conditions required for such a level of connectedness, collaboration and knowledge exchange. Admittedly, we have technologies like Machine Learning, Neural Networks, Natural Language Processing, Cognitive Computing, but there are components that no one is even working on to bring singularity closer. For example, how do you create AI engines that can invent and create new innovations which are contrary to what it has historically learnt from past experiences? I’ll leave you to think about this for a while and maybe discuss it a bit further in the future. For the moment, what I suggest is this: let’s wait for Martin to release Winds of Winter, which is the next part of the GoT epic. We can then compare what the AI has written with the work of Martin. We’ll know who the real master is.
[ii] https://www.youtube.com/watch?v=whv-krWnq0g
[iii] https://spectrum.ieee.org/robotics/medical-robots/would-you-trust-a-robot-surgeon-to-operate-on-you
