Even if it works, you lose control over systems that you don't understand. And that gets expensive.
Ever worked for a bank? Prime example of this. Each of the major banks has one guy that understands the bottom layers of their stacks. He's in his 60s. He hasn't retired because the bank keeps driving dump trucks full of money up to his house. He knows more about banking than the CEO. He knows how to code on punch cards. He's been known to bite. You know a version of this guy, eh?
The current process when you need a change made is to make a request and wait 6 to 8 months and hopefully he gets back to you.
When that guy retires... I dunno man.

If a human can do it, a machine can do it.
Unless you think that abstract thought comes from your "soul" or some other form of mystical garbage.

Ya, all the people that didn't start learning to code when they were young and malleable.

No Alexa, that guy!

The 1/50k number and 1/100k numbers are clearly just made up. The Touch ID figure is a pretty blatant lie, fingerprints aren't even that unique.
And this article is proof that the 1/100k is also bullshit. Siblings for fucks sake!
This reminds me so much of the Nuance voice auth system that was supposedly 1/10k false positive rate, but anybody could log on as anybody else by doing a half-assed impression of their voice.

He meant to say it's like reverse-locusts that show up and barf delicious grain all over farmers fields.

It does seem like whatever device is being used here went wrong in some way. Even if it's some kind of sonic weapon and brain damage was the intended effect, you would think it wouldn't be designed to make a noise.

He's not talking about replacing deep learning, just back-prop. That's the method used for training a network. Hinton thinks that an AI would need to learn without thousands of labeled examples, and back-prop isn't up to the task.
I hope he's wrong, because replacing back-prop would be a real son of a bitch.

I really want to know where Apple is getting these false positive rates from. I've read reports from third-parties claiming that TouchID has a false positive rate as high as 1/200. Fingerprints, man. They aren't THAT unique.
And 1/1000000? First off, thats a suspiciously round number. Also, without a huge specificity for sensitivity trade-off, this just sounds way too good to be true. Kudos to Apple if I'm wrong.

If you're into this kind of thing, check out the Racetams family, or Bromantane. They're the current darlings of the noots community. Legal too (for the time being.)

Oh, my personal favourite Android documentation is for View's onDraw method:

And big surprise, no mention in the style guide of updating documentation when something becomes deprecated.

For a REST API these days documentation has gotten very easy due to the Open API standard. For most major frameworks you don't even need to write any sort of definitions, you just put annotations in your code. Then run it through Swagger UI (or similar) and it generates standardized, informative, and even interactive documentation. It's really beautiful.

The definition of AI has changed. We all need to get used to this. If you throw a fit every time somebody says "AI" without meaning some kind of magical super machine, all of your hair is going to fall out. The meanings of words change over time. Just go with it.

The modern definition is basically just any implementation of machine learning. Which is funny, because the phrase "machine learning" also used to be a buzzword that nobody in the industry actually used unless talking to the media. We'd be more likely to say "Computational Statistics" or "Function Estimation". So, in summary, the modern definition of AI is: statistics on a computer.

In ML that's called "Feature Importance". Some algos can do it. For example, if you're running a decision tree based model such as Gradient Boosting, you can determine which particular pieces of data were considered important in a prediction.
But I'd be willing to bet that's not what Watson is using. IBM is more concerned with making headlines than functional products. So they'll of course go right for the Neural Nets (Works like the human brain!1!) . Which is of course a black box that requires a huge amount of human work to get any idea at all about why it made a decision.

Harder to turn a buck on mushrooms though.