Machine learning on iOS

From Apple's teaser page on the new Core ML feature announced as part of iOS 11:

Core ML lets you integrate a broad variety of machine learning model types into your app. [...] Core ML seamlessly takes advantage of the CPU and GPU to provide maximum performance and efficiency. You can run machine learning models on the device so data doesn't need to leave the device to be analyzed.

A big part of the keynote was privacy, and this is another way to help protect privacy by ensuring data doesn't have to leave the device. But it also seems to be a terrific technical accomplishment - like where Google is heading with Tensorflow on mobile, but here and now with models at your ready.

With Core ML, Apple is adding a ton of capability in image recognition and language processing.

Vision - You can easily build computer vision machine learning features into your app. Supported features include face tracking, face detection, landmarks, text detection, rectangle detection, barcode detection, object tracking, and image registration.
Natural Language Processing - The natural language processing APIs in Foundation use machine learning to deeply understand text using features such as language identification, tokenization, lemmatization, part of speech, and named entity recognition.

These of course join Speech Recognition introduced at last year's WWDC.

Speakers should be for playing music, not shopping

What Hi-Fi reviews Apple's HomePod:

As Sia’s The Greatest played out, the HomePod sounded impressive: strong bass rang out – which was perhaps the overriding audio takeaway for the speaker – but the vocals still seemed sharp and crisp.

In comparison, the Sonos Play:3 appeared uncharacteristically flat, while the Amazon Echo felt almost pedestrian.

I've had a bluetooth speaker for a few years now that's just clunky to get working with my iPhone. I'm really excited about a terrific-sounding speaker that just works. Oh, and having Siri will be pretty neat too.

Via: MacRumors

Why we make plans instead of goals

This is insightful:

You may notice that there are fewer goals at our company than most. Instead of setting goals, we prefer to make plans. Here’s how that works, and how the two relate.
We do start with a goal - a change we’d like to see. A goal is an outcome - it could be a cutting down the time it takes to add content to the app, doubling traffic to our site, or hitting a certain revenue. It’s important to know where you want to get so that you can choose the most important actions every day!
The big problem with goals is that they don’t tell you what to do every day, especially the bigger more all-encompassing goals. If you look at a goal like “get to 20,000 users” - what exactly are you supposed to do next to go get to that goal? Unless your job is sales, you don’t directly affect that goal.
Instead your actions indirectly affect it in various ways. Even if a goal relates to you directly - like increasing traffic to our site if you are in charge of the blog - there are many steps to get there. Just looking at the goal doesn’t help you figure out what to do next, or what to prioritize.
That’s why we don’t make goals, we make plans!
Plans tell us what to do every day. Plans have a clear yes or no whether or not we enacted them. With goals we might fall a bit short, hit them exactly or exceed them - and what does each of those outcomes mean? Which was a success and which was a failure? We don’t agree that we have failed if we did our very best work but the outcome wasn’t what we anticipated. Yes you can say that in the end only the outcome matters - but we believe this is the most effective path to get to that successful outcome.

Via: Software Lead Weekly

An approach to wiping out American poverty

There are growing discussions about the viability of Universal Basic Income as a way to combat poverty and to stimulate new types of economic growth (i.e. startups and small business). But in any of these conversations, we quickly get to the question of how to pay for UBI. The default answer (and a non-starter for most fiscal conservatives) is that it requires some sort of wealth redistribution mechanism such as increased taxes. Which is why this treatment of the topic by Dylan Matthews, writing for Vox, is so interesting.

If you’re not willing to entertain big tax increases, then you should be thinking about a negative income tax. That’s the term used for a universal basic income that tapers off with income. So a negative income tax of $10,000 for adults and $5,000 for children, and a 50 percent phaseout rate, for example, would offer a family of four with $0 in earnings benefits worth $20,000; if they started earning $10,000 in wages, the benefits would fall to $15,000, for a total income of $25,000; by the time they earned $40,000 in wages, they’d be getting no basic income payment at all. A negative income tax is just a UBI financed in part by a somewhat regressive tax on the first chunk of earnings people make, and because of that tax, its net price tag is much lower.

A negative income tax sidesteps the issue with UBI of paying people a basic income who don't need it - like paying Warren Buffet a basic income on top of his massive wealth.

Wiederspan, Rhodes, Shaefer, 2015 In an absolute must-read paper for anyone interested in the basic income debate, the University of Michigan’s Jessica Wiederspan, Elizabeth Rhodes, and Luke Shaefer estimated the cost of the US adopting a negative income tax large enough to wipe out poverty. To be conservative and get a high-end cost estimate, they assume that such a program would discourage work substantially.

Okay, so the cost of people not working is factored into the program. Check.

Despite that, they find that a household-based negative income tax, set at the US poverty line and with a 50 percent phaseout rate, would cost $219 billion a year. That’s almost exactly the same as the combined cost of the earned income tax credit (which supports the working poor), Supplemental Security Income (itself basically a negative income tax but only for the elderly and disabled), food stamps, cash welfare, school meal programs, and housing subsidies. You could swap those programs out, put a guaranteed income in their place, and wipe out poverty entirely.

And there we have it. This analysis shows that if we use the money we're already paying on programs for the poor, we could eradicate poverty in our country. Seems like a good thing to do.

Facial recognition is used to nag you when you're not paying attention to your studies

Amor Too, TechCrunch, reports on software for online courses that uses artificial intelligence and facial analysis to determine whether students are paying attention in class:

The software uses students’ webcams to analyze eye movements and facial expressions and determine whether students are paying attention to a video lecture. It then formulates quizzes based on the content covered during moments of inattentiveness. Professors would also be able to identify moments when students’ attention waned, which could help to improve their teaching, Saucet says.

Even though this is on The Verge, I'm still not sure if this is a joke or not.

Rose Luckin, a professor at the University College London Knowledge Lab, had this to say, “A much more pedagogically sound approach would be to show the student when they are focused, and how that relates to their performance, So that you're offering information back to the student that helps them to structure their work time more effectively and helps them to become a more effective learner.”

Translation: this is a dumb idea.

Blockchain is the new Linux

Jon Evans, TechCrunch, shares a compelling argument:

...blockchains today aren’t like the Internet in 1996; they’re more like Linux in 1996. That is in no way a dig — but, if true, it’s something of a death knell for those who hope to profit from mainstream usage of blockchain apps and protocols.

Decentralized blockchain solutions are vastly more democratic, and more technically compelling, than the hermetically-sealed, walled-garden, Stack-ruled Internet of today. Similarly, open-source Linux was vastly more democratic, and more technically compelling, than the Microsoft and Apple OSes which ruled computing at the time. But nobody used it except a tiny coterie of hackers. It was too clunky; too complicated; too counterintuitive; required jumping through too many hoops — and Linux’s dirty secret was that the mainstream solutions were, in fact, actually fine, for most people.

We're going to see fantastic creations because of blockchain; and perhaps blockchain will fundamentally change the financial industry. But blockchain isn't the revolutionary change the the internet was (and continues to be). The Linux analogy is a better fit.

At present, there is no practical means for testing the safety of AVs prior to widespread use...

According to the Bureau of Transportation Statistics, there were about 35,000 fatalities and over 2.4m injuries on American roads in 2015. While these numbers sound high, given that Americans drive three trillion miles a year, the accident rates are remarkably low—1.12 deaths and 76 injuries per 100m miles. Because accidents are so rare (compared with miles traveled), autonomous vehicles “would have to driven hundreds of millions of miles, and sometimes hundreds of billions of miles, to demonstrate their reliability in terms of fatalities and injuries,” says Nidhi Kalra of RAND Corporation, a think tank in California. At present, there is no practical means for testing the safety of AVs prior to widespread use. For many, that is a scary thought.

That's a lot of miles. Also, the article has good examples for each of the different levels (Level 1-5) characterizing the capabilities of autonomous vehicles.

Via The Economist