Most startups cater to every new kink the investors develop. We've all seen them trying to ram machine learning shaped pegs into their product shaped holes. Who could blame them? Capitalism, for better or worse, ensures that the ones that don't bend to the whims of the market are replaced by the ones that do.
We were having dinner after racquetball, and the topic turned to machine learning. One of my friends said that, and I'm wildly paraphrasing here, it is the garnishing, not the secret sauce. Couple of us disagreed. He asked us to name a few companies that have built successful products using machine learning as the core tech. We mumbled umm... Sift Science, Waymo... We couldn't come up with many.
His argument essentially is that machine learning can be used to improve existing products, not create novel ones. Or at least that, most products that claim to use it won't have a machine learning heart, but will be wearing a machine learning lipstick. It is an optimizer, and hence it could only be used to improve things.
My treppenwitz argument is that everything is an optimization problem.
While it may be true that machine learning is mostly being used to enhance existing products or getting investors to open up their wallets, I think we'll see a lot of truly novel applications soon.
I was turning it over in my head, trying to express exactly why, when I stumbled on this beautiful article by Andrej Karpathy himself, which captures most of my thoughts on this and expresses the points succinctly: Software 2.0 (mirror)
Instead of spelling out what to do, you loosely describe what you want, and let it figure out how to get there. This is a new way of writing software. That's the most exciting part for me.
The amount of things I need to learn to get anywhere marginally good at this is simultaneously incredibly thrilling and extremely daunting.
Thanks Alan, for suggesting improvements.