Jump to content

The Sorry State of AI


Recommended Posts

This is really an excellent article, that should be mandatory reading for anyone hyping and hoping for big AI breakthroughs just behind the corner (perpetually "any moment now"). A good not too technical description of the promise and the limitations:




"Just about every AI advance you’ve heard of depends on a breakthrough that’s three decades old. Keeping up the pace of progress will require confronting AI’s serious limitations."


Spot on. This is why I have always called current AI just a glorified expert system, a bigger pocket calculator, "dumb" algorithm-driven data mining. As the article points out, there is work that's going on AI that attempts to transcend this primitive "bigger pocket calculator" one trick pony limitation, but as one of the AI giants observed, this will take *generations* to crack. My belief exactly. I think we are decades and decades - at a minimum - from something that is worthy of the name of true AI, and more likely centuries (200 years?). This was an area of interest to me back in the early 80's at the uni, and from a cursory reading of what AI is up to today, there's been virtually no progress whatsoever on true AI (though a lot of data mining thanks to cheaper computational devices). 

Link to comment
Share on other sites

You're so 2012.



Baby steps:


More storage and processor speed, and you will have a pretty convincing artificial toddler that can walk and talk.  The pace of progress isn't lightning fast, but its moving...


Paraphrased quote from a comment on your article:

"The question about why a child can learn more fluidly than a computer today is absurd. Of course she can, her neural abilities are the product of millions of years of refining, evolving, automated subprocessing. Neurotypical people compute facial recognition easily.. folks on the autistic spectrum don't, because subprocessing in that particular area is impacted. And that's just one of so many unconscious subprocesses that are vital to us as a social species. We're only a couple decades into the process of trying to recreate that evolution.  Many are more interested in producing things that are currently useful to us and handle the 'slog' factor (things that are tedious and time consuming) than reproducing the human intellect. If we really wanted to, we'd start with engineering a fish (or mouse) brain and take it from there."
Link to comment
Share on other sites

Tell you what. Here's a real-life test to settle this argument. I will admit that AI has reached some kind of landmark, if it can do the following - which has already been promised/thretened as coming about within practically months:


Fully driverless cars could be months away




If I can order a car to drive - 100% without a driver on board, and without any human remotely steering it - from NY to LA completely autonomously. The only human intervention should be to put gas in the tank at given intervals. 


My claim: this will not happen for decades and decades - not within our lifetimes. 


For that to happen, you'd need an order of intelligence and infobases and data processing that's going to be out of reach of AI for a long, long time. The reason is extremely simple - there are whole classes of unanticipated possible events that might happen on the road that a human can correctly react to based on a vast array of disparate disciplines. AI as presently structured, can only react to very limited worlds with extremely tight parameters - and so you can have cars drive along special race tracks, or roads that have been extensively annotated, counting on a limited number of events. But what about a situation that human drivers deal with extremely easily on almost a daily basis - go from point A to point B along roads you have never driven before? AI cannot do this. Today, all AI can do is go along roads that have been exhaustively documented and datamined - that's not how human AI deals with the challenge, and therefore it is a challenge that true AI would have to meet and which it fails at abysmally.


I think this is a very fair challenge and not asking for the moon - after all, we're told practically daily how truly driverless cars are just around the corner. So, let's see!


That's the challenge. My prediction is that this challenge will not be met within our lifetimes. 

Link to comment
Share on other sites

I tend to to agree. When I was young 50 years ago we certainly were on the verge of curing cancer, the common cold, ending world hunger etc. Now its aging, global warming and yeah driverless cars etc. I'm always hopeful but not holding my breath.

That's because we reached the moon, cured polio, smallpox, so people thought everything is possible. But we were extremely ignorant back then. Reaching the moon is not nearly at the same level of complexity as making flying cars for everyone.. Just like curing polio with a simple vaccine is not even remotely comparable to curing cancer. And yet people thought so. Scientists simply had insufficient knowledge back then. Only when you have sufficient grasp of the subject, you can make more reliable predictions.


Today we understand cancer fairly well and not many are claiming it will be eradicated any time soon. The cure will come gradually (just like driverless cars) and finally materialize sometime in the second half of this century (probably, I guess, I hope).

Link to comment
Share on other sites

Without a lot more details, it's hard to say what this means. Has the rout been exhaustively documented before? Is this the very first time that the truck drove this route, without trial runs first (i.e. what a human driver would be capable of)? Was there some kind of remote control dial up from the truck where it could consult human operators in case it was confused by some circumstance or another (how many self-driving vehicles actually operate). And so on.

Link to comment
Share on other sites


This topic is now archived and is closed to further replies.

  • Create New...