I used a Roomba vacuum cleaner for the first time recently while staying at a friend’s house. A couple of friends had raved about it: “all you have to do is just press a button and it does the vacuuming for you.” It is hard enough to find a human that can clean to my standard, so I was skeptical that a 12 inch black plastic moving circle that is not even as tall as a cat would clean to a satisfactory level. I picked a simple rectangular room with minimum furniture to start with. I pressed the clean button, the cat and I watched it set off, moving back and forth, diagonally. The cat lost interest in about two minutes so I was left to observe by myself. I quickly noticed that it seems to be going round doing the same spot over and over again but not able to pick up the debris. I thought maybe I should choose the “spot clean” option. I did but it still struggled. Like the cat, I gave up! I turned to the older human assisted vacuum cleaner and did the room in a couple of minutes, while the Roomba couldn’t finish half the room in 15 minutes. This experience made me think of questions I have about AI (artificial intelligence) replacing humans in many spheres as it is touted to do.
AI Will Do Away With Humans There is no shortage of headlines that say: “AI can solve climate change”; “AI to do away with doctors”; “AI discovers new antibiotics”. Tech companies also love to hyperbolize their AI capability and how it is so much better than humans. For example, in 2016 DeepMind announced its AI program AlphaGo beat Lee Sedol, the world’s strongest Go player, at a game. Go is supposed to be much more complicated than chess so this got a lot of attention. What got little attention is that it took 100+ top scientists and engineers and an army of super computers to beat Sedol. So why promote the ‘AI has surpassed and will replace humans’ narrative? Why is even a goal to replace humans with AI? My super-duper software engineer friend says it is because humans are expensive compared to computers. Okay, how can one Go player whose winning earnings is around $1million be more expensive than over a 100 scientists working for DeepMind who I have no doubt make high 6 figures? Then there is the cost of hardware for building and running these algorithms. It could be argued that the machines are new to playing Go while humans have been at it for hundreds of years, and the cost of building hardware will get cheaper. Perhaps! It is hard to know for certain how much it would cost to run AI models in the long-term. We still don’t understand the full cost of building and training AI models. A recent paper from researchers at the University of Massachusetts, Amherst, has shown for the first time the environmental cost of training a typical AI model. The study found that training, at baseline, an AI model can emit nearly five times the lifetime emissions of the average American car (and that includes manufacture of the car itself). Don’t get me wrong, I appreciate some things that AI has done for me: I used to have to talk to strangers to ask for directions when lost, now I don’t have to thanks to Google Maps; I buy more items than I need thanks to Amazon’s friendly recommendations. I also have a long list of things I would LOVE AI to do for me: power a washing machine that not only washes my clothes, but folds them and puts them away too; To do ALL my paperwork for me; make trains in London actually run on time; Oh and sort through the maze (and haze) that is my head, pick out words to write the masterpieces I dream of writing. Judging by a recent account of the writing produced by OpenAI’s GPT-2 writing AI (supposedly the most advanced around) program in the New Yorker magazine, I still have years and years to wait for this to happen. The question is, how long do I have to wait? When will AI do away with all human jobs so I can enjoy a life of leisure, living on universal basic income given to me by my government? I also wonder how can we create intelligence that is human level and even superior when we don’t fully understand how our brain works? For answers, I turn to far more learned folks than me. I’m fortunate enough that my job brings me into contact with scientists and engineers that work in the area of AI so I know how complex and hard it is. I also rely on the following sources: Credible Sources Rodney Brooks’ blog: Brooks, an ex-director of MIT’s Computer Science and AI Lab, a robotics entrepreneur and author offers a no-BS perspective on what AI really is (as opposed to the myriad of labels attached to it) and what it can and can’t do. Incidentally, Brooks is also a founder and former CTO of iRobot, the firm that created the Roomba, so he is very familiar with turning AI into real products. His blog is full of gems. A good place to start is his Seven Deadly Sins Of Predicting The Future Of AI. Since 2018 he has published a list of predictions on what and when things will happen in AI in the next 30 years. He updates the list every year. Highly recommended. Lex Fridman’s AI podcast: Fridman is a researcher in human centered AI, deep learning and autonomous vehicles at MIT. This podcast has really improved the quality of my work-out sessions, I feel like I burn double the amount of calories from the learning I get listening to it! He asks thoughtful and probing questions of his guests from the fields of AI, maths and cognitive science, which makes for deep conversations that are still accessible to layman like me. This is currently my favorite podcast. Check out the recent episodes I’ve enjoyed: Ayana Howard: Human-Robot Interaction; Daniel Kahneman: Thinking Fast and Slow: Deep Learning and AI; Melanie Mitchell: Concepts, Analogies, Common Sense and Future of AI
The Algorithm Newsletter from MIT’s Techreview’s reporter Karen Hao. I get all the latest credible news on AI from here.
It may seem like I favor MIT sources, I don’t, these are just the best I’ve found for my needs. I’m always happy to check out new recommendations so if you have any please send them.