Machine Intelligence Progress is Accelerating

I don’t (yet) worry about a Terminator or Matrix like scenario where machines take over the world and subjugate humans. But it is useful to remind oneself of the rate of progress that has been taking place. I started programming computers in my early teens which is now almost 35 years ago. During that the first 25 of those years progress towards any kind of artificial intelligence was painfully slow. For instance, as late as 2004 when DARPA ran the first Grand Challenge it looked like we were a long way from a self-driving car – the best vehicle made it barely 8 miles into a closed course in the desert. But then development took off and by 2012 only 8 years later Google driverless cars had covered several hundred thousand mies on public roads.

Other areas have seen similar acceleration. For instance, image understanding tasks such as facial recognition used to be notoriously difficult. But in 2014 an algorithm for the first time outperformed humans on a large data set of faces. Further advances in image analysis are coming from the progress that has been made with deep learning. This refers to a set of technologies that allow for the formation of intermediate concepts based on observed data. That turns out to be closer to how the human brain works and is producing some spectacular results. For an interesting demo of the latest capabilities you can check out the recognition capabilities at Clarifai, a New York based startup. Again the rate of change here in the last few years completely outpaces anything we have seen previously.

And here is yet one more example of progress. Realistic walking was considered to be one of the things robots would have a very hard time with. And we are still not there, but compare this video of an early Big Dog from Boston Dynamics (now owned by Google) with the latest generation called Spot

Now one might look at these in isolation and see only highly specialized somewhat awkward machines that can’t possibly have much of an impact on say the labor market. But that would be an error. Machines can and will do things quite differently and that is what will let them outperform. And no single machine needs to outperform humans at all tasks. Instead all it takes is for specialized machines to do so on the tasks they were designed for.

That is why I will keep writing here on Continuations about such things as Basic Income Guarantee. Having followed the field of machine intelligence for many years I am excited by what now lies within our reach and want us to push forward. We just should simultaneously work on all the public policies that will be needed.