How Google Is Remaking Itself as a “Machine Learning First” Company

>> Traditional AI methods of language understanding depended on embedding rules of language into a system, but in this project, as with all modern machine learning, the system was fed enough data to learn on its own, just as a child would. “I didn’t learn to talk from a linguist, I learned to talk from hearing other people talk,” says Corrado.

It's becoming more and more important to me to make sure this sort of insidious fudging of the truth with the ultimate goal of overhyping Google's tech does not go unchallenged. Why is it important? Because leaving this nonsense unchallenged is basically leaving Google unchallenged to dominate the market- and we end up with crap tech, like we did when Intel's crap CPUs dominated.

(yeah, x86 is a piece of crap- the Von Neumann bottleneck and so on)

So here's the problem with that paragraph above: yes, children don't learn language by studying its rules. However, they also don't learn it by sifting through terrabytes of data and optimising the parameters of a model in their heads.

Hell- adults don't learn like that. I live in a foreign country where I learn new words and expressions daily. I am still amazed at my own ability - no, my own _brain's_ ability- to learn how to use such new words and expressions by just hearing them once in context. It's like the way to express old meanings in new ways, or even whole new meanings, somehow nails itself firm into the middle of my old... context? Or whatever you might call it. It's just finding a new piece of a puzzle behind the sofa. Somehow, the moment I click it into place, it fits in perfectly and stays there for good. I don't need to repeat it a thousand times to myself or see it in the same place in a thousand puzzles. I just... learn.

Obviously I have no idea what really goes on inside my head when I learn a new language, or whatever goes inside childrens' heads when they learn their first language, or what went on in my own when I did.

But one thing is exceedingly clear: human learning is _nothing_ like machine learning.

You want to argue that machine learning is better than hand-crafting rules for a rule-based system? Go right ahead. There's arguments for and againsnt and there's a healthy debate to be had.

But don't argue that machine learning is the way to go because "it's how people learn".

Because it's not. It's nothing like that.