Mercilessly pricking the bubbles of AI, Big Data, machine learning


#1

Originally published at: http://boingboing.net/2014/10/21/mercilessly-pricking-the-bubbl.html


#2

I have a hard time believing very many of the people who are working in AI day in and day out are within the bubbles Jordan is popping. Anyone who thinks AI can be “solved” is pretty delusional. That’s like saying we’ve solved Chemistry.


#3

It’s not that people who actually work in the field believe in AI and Big Data fantasies, it’s tech bloggers and journalists and PR people and analysts who profit from inflating trends into “actionable” investments from CIOs.

Big Data is just a minor hype-cycle blip on the radar of the major analysis firms like Gartner that profit from promoting these things year in and year out. Some time ago Google had to solve a particular hard problem with a nontraditional database. They did so, and publicized their approach, and as usual the world decided it was the solution to everything under the sun. Of course the vast majority of so-called Big Data applications can actually be implemented in small conventional databases or in spreadsheets, and vanishingly few require a distributed map-reduce framework for whatever computation or queries run over the data, but that concerns most of the consultants promoting Big Data solutions not at all.

Similarly there really have not been any significant leaps forward in AI. Not ever. Progress is very slow and incremental. Indeed much of what is fondly called AI these days is some variation on emergent-data analytics, by no great coincidence another Google specialty. Yes, yes: I give the Watson team all the credit in the world, but that was a very specifically narrow application area, and it really hasn’t translated into anything worthwhile outside of publicity for IBM (who needs it desperately given their current financials).


#4

I’m similarly skeptical of anyone who thinks that AI will naturally gravitate toward creating a human-like intelligence. That’s like saying the aerospace field is gradually progressing toward a really good artificial bird.


#5

While I respect what Michael Jordan has done for the game, and I know he has had direct experience with Dennis Rodman; I’m skeptical that he could speak with much authority on the subject of AI in the academic sense.


#6

The AI hegemony will have a special place for people such as Michael Jordan.


#7

Wintermute is on the case.


#8

The AI is shedding fields left and right. Speech recognition used to be AI, a lot of machine vision as well… tempts me to say that AI is defined as what computers cannot do yet.


#9

I think that is kind of why he is writing this. Typically people in a field are skeptical of grandiose claims because they know these will eventually lead to a public backlash. The problem is that sometimes such skepticism gets distorted by blogs and other media to say “famous person X thinks his field is rubbish”, leading to a premature public backlash. This sort of happened in genomics where several famous genomicists stressed in the wake of the Human Genome Project that the genome wasn’t enough to understand human biology. Typically that got distorted to mean “Genomicists admit that the Human Genome Project was a waste of time”.


#10

Yes, the obvious joke. He sometimes presents at biology conferences (because some of his machine learning work applies to biological data), and the people introducing him never fail to make a Michael Jordan joke as if they were the first person to think of it. I feel for him.


#11

I agree. My comment was directed at the way Cory Doctorow presented this interview. Doctorow is doing exactly what you describe which is funny since Jordan, as you state, is trying to address the sensationalism of science reporting. As someone with a history in the field of AI the phrase “mercilessly pricking the bubbles of AI” jumped out at me. However I was glad to find that the headline was pretty misleading.


#12

Like the relationship between philosophy and science!

So to summarize:

For any given hypothesis, more data is better. But if you try all possible hypotheses against your data, you’ll drown in computational intractability liberally peppered with false positives.


#13

The thing I miss on philosophy to take it seriously is a mechanism to weed the crap from the treasure, a way to kill unviable hypotheses. Once you add this, it ends up as science.

A possibly apocryphal story, of a math-physics faculty head, complaining.
“You physicists always ask for expensive gear. Look at the mathematicians, all they need is a pen, a paper, and a waste paper basket. Or look over there at the philosophers, they need just a pen and paper.”


#14

Agreed.

And some AI researchers whose blogs I’ve read see themselves as reducing epistemology, metaphysics, and morality to AI, self-modifying AI, and friendly AI, respectively. Basically, after millennia of debate we actually need philosophers to be rigorous enough to make a serious dent on those problems, with extremely high long term stakes for getting it right or wrong.


#15

In other words, philosophers have to become scientists.


#16

Every branch of science was part of philosophy at one time.


#17

In my experience AI is best understood as a marketing term. I have yet to meet anyone in that field who actually buys into the Frankenstein bullshit usually associated with it in popular science coverage or fiction. However to non-experts “Institute for Artificial Intelligence” sounds so much nicer than e.g. “Institute for Really Versatile Algorithms”.


#18

I don’t think I was out of my teens when I figured out that AI is purest PR. Oh, and don’t forget “expert systems.” I would probably be a richer man if I didn’t have such a poor attitude towards buzzwords.


#19

In regards to neuromorphic hardware:

“It’s really just a piece of architecture with the hope that someday people will discover algorithms that are useful for it. And there’s no clear reason that hope should be borne out.”

What. My lab has built a ton of useful stuff (hierarchical reinforcement learning, language comprehension) with the NEF and it can run on neuromorphic hardware.

If you want more details on the cool shit we’re making check out this:
http://1.usa.gov/1tg2n2I


#20

There is nothing to indicate that spending all that money on the HGP was more productive than spending the same amount on hypothesis driven science. Indeed the HGP was the quintessential big data project with bad quality control and bad data cleaning, as well as a dubious statistical model (Remember how trio data was supposed to become obsolete?).

At the end of the day “Big Data” is driven by the same market forces as most IT business fads - persuading companies to spend more money on servers. This is the same thing that drove the HGP - spending billions on hardware and laying off the staff based on the assumtion that with the right hardware magic just happens.

And again the magic of Big Data plays into the vanity of developers who have a personal vision of an API that will make SQL obsolete using a half million lines of OO code. There should be some sort of monument to all those projects.