17 October, 2017

Machine Learning, Big Data and Artificial Intelligence

Listened to a really good episode of the TED Radio hour podcast on my way into the office the other day. Clued me in to a TED talk by Sam Harris (he's written a number of books that were New York Times best-sellers and is also a credited as a neuroscientist and "armchair philosopher") - basically the talk is about the (as he sees things) inevitability of a super AI evolving (computers are already routinely teaching themselves to learn to do new tasks without human assistance and doing better than their original human programmers), however it tends to be just one thing at a time. When they do learn things they tend to learn to do them very well - play "Go" or chess for example - computers are unquestionably better than humans now at playing these games.

Cars are rapidly learning to drive themselves - people have been involved, but largely the computers are using trial and error in simulations like Grand Theft Auto (where they can't kill anyone real) to learn to drive around complex 3D environments without having a crash. Eventually (and soon) most computer scientists agree computer AI's are going to be doing a lot of the work humans do, are going to be able to perform research unimaginable a few years ago (once the internet of things really happens - already my 'fridge at home is connected to internet) - computers are going to start solving problems humans didn't even know to look for. Looking through millions of variables and tie together say (I pick a few random things here) - the number of times I'm opening the 'fridge making the temperature fluctuate, the fact I have eggs in there from a certain farm, data from an automatic stool/urine sample taken when I visited the bathroom, data on how aggressively I apply the break in the car when driving (if I'm still allowed) and find a pattern of behaviour/things that reliably predicts (for example) I'm in the early stages of developing a brain tumour. Imagine what an AI is going to be able to do once it has all the data on the planet and just starts rapidly looking for patterns -

Things unpredictable and previously non-testable can all come together once we have this huge matrix. Anyway, I digress into the (equally fascinating) subject of big data, the talk is really about making sure we don't do a couple of things (and why) -

  • Get into an uncontrollable arms race between competing AIs (President Putin and Elon Musk agree - I'm sure scientists over Russia are working away on the problem as I type and Google are open about their developments) 
  • Allow ourselves to unwittingly settle into the position of the ant, if the AI is put in the positional equivalent of a team of construction workers - normally not a lot happens, construction workers don't tend to hunt down ants they just live alongside them, but if we have even a slightly differing priority at any point the AI just runs all over humanity without a second thought like a busy guy with a job to do sat in a big JCB 
  • Ignore what we are getting ourselves into - effectively - compared to the sum of current human ability - we are busy setting about building ourselves a god - and to use Sam Harris' words "the train has already left the station" - it's coming and it will be real within the next 50 years.

Very interesting stuff, you can watch the video here - What happens with humans develop super intelligent AI - or listen below -

No comments: