[thr0waway1239]

TLDR from article which supplement's eli_gottlieb's comment:

"Much of today’s A.I. boom goes back to 2006, when Amazon started selling cheap computing over the internet. Those measures built the public clouds of Amazon, Google, IBM and Microsoft, among others. That same year, Google and Yahoo released statistical methods for dealing with the unruly data of human behavior. In 2007, Apple released the first iPhone, a device that began a boom in unruly-data collection everywhere."

The combination of smartphone + cloud created a virtuous (for AI, that is) cycle of unending data collection, which fed the improvements in theoretical research models which needed data at a larger scale to be validated.

[api]

I remember doing AI work in 2005-2007 (genetic programming and similar things) and one of the hardest thing back then was that good quality training data sets were really hard to come by. I can second the idea that data makes a huge difference, and there's tons of it today.

[vonnik]

AI is moving fast. Faster than it has in a long time. In research circles, "supervised learning", or data classification problems, are considered solved. Consider the magnitude of that statement: If we have sufficient labeled data, we can predict those labels accurately on data we haven't been exposed to: fraud, faces, diseases, you name it.

AI is moving fast for three reasons, which Andrew Ng summarizes neatly: 1) We have more data than ever before (some of which has been organized in fabulous datasets like ImageNet), much of which is being generated online or by sensors. 2) We have more powerful hardware than ever before -- this is the combination of distributed computing with GPUs. You could say that AI advances at the speed of hardware, or at least is limited by hardware advances. Brute force AI like deep learning is the main beneficiary. 3) We have seen a steady stream of algorithmic advances for years. Specialists despair of keeping up with the literature, research in AI is so feverish.

Others have pointed out additional factors: open-source projects that lower the barrier to entry for the algorithms; cloud computing services that open up access to the hardware. The choke right now, and what's slowing wider adoption, is skills. Many companies don't have the teams to implement AI well.

Some of those companies are also the noisiest about selling AI. I have deep doubts about some vendors to deliver on the hype they're encouraging. And I believe that will lead, not to an AI winter, but to a poisoning of the well that the AI sector will have to address for years to come.

But overselling and hype are inevitable symptoms of real advances in tech, and in some ways, the tech is advancing faster than the hype can keep up with, precisely because the people hyping it, whether salesmen or journalists, don't understand the true extent of what's going on.

AI is just math and code. In a sense, you could say we've entered the age of Big Math, which is the next stage after Big Data. The math is necessary to process the data and determine its meaning. The math takes the form of massive matrix operations that can be processed on the parallel calculators known as GPUs. To call it statistics, as the reporter of the piece does, is only partially true. The math involved in AI comes from probability, calculus, linear algebra and signal processing. It's more than fancy linear regression. And it's definitely more than just making predictions about customer behavior, however attractive that is for industry.

Asking why now about AI software is like asking why now about cars after the Model A came out. Because it's there, it's faster than horses, and it makes you look cool. Like cars or any other powerful technology, AI is part of a race, and that race is taking place between nations and companies. You can decide not to adopt AI, the same way newspapers decided to ignore the Internet, or the way the French decided to fight German panzers with mounted cavalry in WWI. There really is not choice whether or not to adopt AI-driven software. It's a question of when, not if. And for many companies, the when is now, because later will be too late.

To give one example of how fast it's moving: Deep learning has been widely thought to be uninterpretable, or without explanatory power, but that is changing with cool projects like LIME: https://homes.cs.washington.edu/~marcotcr/blog/lime/ Which is to say, for some problem sets, we'll be able to combine the impressive accuracy of DNNs with the reasons why they reached a given decision about the data.

On the hardware front, NVIDIA and Intel are racing to build faster and faster chips, even as startups like Wave Computing or Cerberas come out with their own, possibly faster chips, and Google works on TPUs for inference.

[apsec112]

"the way the French decided to fight German panzers with mounted cavalry in WWI"

Quick note: I think this is historically inaccurate. In World War One, the Germans had essentially no tanks, unlike the British and French. In World War II, there's a popular rumor that Polish (not French) cavalry forces charged German tanks at Krojanty, but this is itself a myth (the cavalry attack was against infantry, tanks arrived afterwards).

[YeGoblynQueenne]

>> In World War II, there's a popular rumor that Polish (not French) cavalry forces charged German tanks at Krojanty, but this is itself a myth (the cavalry attack was against infantry, tanks arrived afterwards).

Not just any cavalry- Polish winged hussars:

https://en.wikipedia.org/wiki/Polish_hussars

Obviously, they could totally take a tank single-handedy, but they have now ascended into heaven to party with the Valkyries.

(It is known.)

[lifeisstillgood]

I did not know that about the cavalry charge occurring before tanks arrived (it is a little suicidal so is surprising)

[vonnik]

Sorry, that should have been WWII, you're right. Anyone who wishes to know more about France's lack of preparation for both wars should read the French historian Marc Bloch, who fought in both and died in the resistance. L'Etrange defaite is his great work.

[js8]

> It's more than fancy linear regression.

I am not an expert, but it seems to me that "big data algorithms" are actually simpler and require less ingenuity than classical statistics. Just look at a difference between frequentist and bayesian approaches in statistics. In frequentist methods, you need to choose model wisely, you're constrained what you can model, and there are many different models with different assumptions. With bayesian methods, you just represent any distribution somehow and do the calculation on those. Less assumptions for the models, more general models, but at the expense of more computations.