We'll predict your EVERY MOVE! Facebook's new AI brain talks to El Reg

A minor flaw in your analysis

> "assumption that humans can be modelled in reductionist materialist terms"

To have some success, the AI engine, for want of a better term, only needs to do some pattern matching to make some conclusions. That's how spam filters work or http://www.forbes.com/sites/kashmirhill/2012/02/16/how-target-figured-out-a-teen-girl-was-pregnant-before-her-father-did/

They don't have to model everything, nor do they have to figure out everything to be very useful to an advertising machine like Google or FB.

The spooks might use this, but they are less motivated to pry into your life. The people who ARE very motivated, and are already doing this to you, are the advertising engines.

You are not Facebook's customer. You are Facebook's product. They sell you to the merchandisers.

Re: A minor flaw in all this

I thought all AI so far was a flop so far

You thought wrong. AI technology has been hugely successful, and profitable, in many industries. Natural Language Processing alone is a major software segment, covering things like advanced indexing and retrieval, sentiment analysis, automated translation, voice recognition, etc.

Facebook and Googles weak AI ... all are built on the assumption that humans can be modelled in reductionist materialist terms

Wrong again. So-called "weak" AI is specifically that part of AI research that doesn't aim to create a sentient machine.

For that matter, even strong AI doesn't necessarily aim to model, or reproduce, human cognition; or when it does aim for human-like cognition, doesn't necessarily claim that human cognition per se is mechanical; or when it is predicated on the thesis that cognition is mechanical, it does not necessarily assume that - there are quite a few complex arguments supporting that thesis.¹

This is one of the reasons why many researchers working in machine learning stay away from the AI label - people who know little or nothing of the subject love to pronounce on it.

And, incidentally, anyone working on strong AI, or even weak-AI projects that aim to predict human behavior, is aware that people don't always do what a model expects. That's kind of, y'know, axiomatic. Even economics has finally seen the light on that one (with the burgeoning field of behavioral economics).

¹For example, John Searle, of "Chinese Room" fame, is often cited as arguing against the possibility of AI, but in fact he explicitly stated that he believed thinking is a mechanical process and strong AI is possible. He just thought a certain class of approach ("symbolic manipulation") couldn't accomplish it.

I believe that is their biggest hurdle.

When giving an advanced AI the task of spending all its time on FaceBook, how to keep it from deleting itself!

Re: And so we take one more step

Oh good lord. You'd think no one ever built a machine-learning system before to model user behavior. People have been doing this for half a century now.

Re: Subtitle is....

I signed up for all three. Not with my real identities of course! I'm not a complete zonk!

Only use twitter for work (make work-related announcements to students).

Left FB from general boredom with it.

Never officially closed my SL account, but only use OSS VR environments these days (and not for social things, generally).

Re: Subtitle is....

I can't tell you anything about FB or twatter, but I know that SecondLife still has a large userbase. And bonus is land is really cheap now since all the people with actual lives left.....

But for those of us with a fat pipe to the house but a 50 km drive to the nearest village, it beats the price of petrol.

Re: Cool. Facebook can continue carrying on indefinitely without our participation.

Does anyone with a cerebral cortex believe this bullshit?

To what bullshit do you refer?

Facebook's announced plans to model user activity? Perfectly doable and quite probably profitable. Given the size of their user base and transaction rate, even just a whacking great Hidden Markov Model would probably correlate pretty well to a lot of the activity. Stacked neural networks (which is what Lecun apparently specializes in) should do even better.¹ They (unlike the Reg's headline writers) aren't claiming they'll correctly predict every action of every user. Getting it right a decent fraction of the time is enough to monetize (as higher-tier advertising services).

If you're talking about the comments to this article - yeah, they're mostly bullshit. Unfortunately many people do seem to believe them.

¹Note that the "deep learning" recently championed by Google is also basically just a stack of neural networks, where higher-level networks make predictions based on the outputs of lower-level ones. The idea is that higher-level networks model abstractions of the lower levels. This seems to be the hip technique this year. Next year it might be P-algebras (generalizations of k-nearest-neighbor) - I'm seeing some noise about that.