Without Humans, Artificial Intelligence Is Still Pretty Stupid (wsj.com) 96
Christopher Mims, writing for WSJ: The internet giants that tout their AI bona fides have tried to make their algorithms as human-free as possible, and that's been a problem. It has become increasingly apparent over the past year that building systems without humans "in the loop" -- especially in the case of Facebook and the ads it linked to 470 "inauthentic" Russian-backed accounts -- can lead to disastrous outcomes, as actual human brains figure out how to exploit them. Whether it's winning at games like Go or keeping watch for Russian influence operations, the best AI-powered systems require humans to play an active role in their creation, tending and operation (Editor's note: the link could be paywalled; alternative source). Facebook, of course, is now a prime example of this trend. The company recently announced it would add 10,000 content moderators to the 10,000 it already employs -- a hiring surge that will impact its future profitability, said Chief Executive Mark Zuckerberg.
Out of date article^W summary (Score:1, Informative)
FTFS:
Yet last month we heard that human-in-the-loop AI is actually inferior to human-less AI for Go. [slashdot.org]
Re:Out of date article^W summary (Score:5, Insightful)
Re: (Score:1)
You're arguing about whether playing Go is AI or not. It's not adding anything, since you (or people who thinks like you) post the same response anytime anyone talks about AI.
It's nonresponsive since I was talking about how the summary used Go as an example of needing humans to help train the NN, and the /. article from last month was about how the NN trained itself without the humans just fine.
Re: (Score:2)
Except the humans invented the game, mades the rules, the computers, the software, the electricity to power all that., etc. Without the humans? not so much.
True, the computer reinvented a better way how to play Go, one that humans likely cannot emulate. That in itself is interesting.
Re: (Score:2)
Without the humans? not so much.
The word "Artificial" means "made by humans". So it seems pretty silly to say it isn't Artificial Intelligence if humans are involved in making it.
True, the computer reinvented a better way how to play Go, one that humans likely cannot emulate.
Wrong. Ke Jie [wikipedia.org], the world's strongest player, lost 0-3 to AlphaGo. He then studied AlphaGo's tactics, and went on a 22-game winning streak against other human players. Statistically he should have won only 60% of those games. So in this case, humans did not teach AI, AI taught humans.
Re: (Score:2)
*slow clap for humanity*
Yes, in a discussion about how to play Go and how computers learn to play Go somehow who created computers, who invented Go, and who paid the powerbill was somehow considered an irrrelevent detail. The point is, the article said (paraphrased) "AI (read as NN) cannot learn without humans in the loop in applications from Go to...." Which is wrong, because less than a month ago was an article about how AI (read as NN) without a human in the loop was superior to AI (read as NN) with a
Re:Out of date article^W summary (Score:5, Insightful)
As long as you live in the fantasy that human brains are magical and computers are not there's no point in having an argument, because you've defined the answer "Humans are intelligent and computers are not, hence anything done by computers is not proof of intelligence." rather than the question: "What is intelligence?" and failed to make any measurable definition or criteria. It's like saying humans have souls and rocks don't and expect the debate to be anything other than theology and philosophy.
Even classical conditioning like Pavlov's dogs are proof of learning and reasoning, hear a bell often enough when you're fed and you associate the bell with food. A plant can't be conditioned, snip off the branches stretching towards the sun and they'll just stretch again and again. If you call it AI and it's not learning it's not really intelligence at all really, if you've found a flaw in a game's "AI" and it keeps falling in the same trap over and over it's just blindly executing. The neural nets at least got that part right, walking into a trap will assign that action negative weights. That's above zero intelligence.
Re: (Score:2)
Although I'm completely with you on about everything in your post, it should nonetheless be noted that even plants are capable of pavlovian learning. Or so it seems: http://rdcu.be/yuI0 [rdcu.be]
Re: (Score:2)
Re: (Score:2)
Something like self driving that doesn't just stop when a truck is backing towards it.
Stopping was fine. Honking would have been smart. But if a little old lady had tripped over and couldn't get up instead that truck driver would have backed over her too, clearly he had no control over what was behind him. So if you're asking whose behavior was the stupidest, man or machine I'd go with the truck driver.
Re: (Score:2)
Re: (Score:2)
So intelligence is something that isn't developed by humans until about the age of 3 - 5? Babies are equivalent to a rock in terms of intelligence?
"Working with an enormous or possibly infinite rule set is what I would call real intelligence."
And therein lies the problem. What YOU call "real intelligence" doesn't matter, it's a classic example of the no true Scotsman fallacy.
The fact is, intelligence is incredibly hard to define, and in reality it exists on a spectrum. Some dogs are dumb and can't even sit,
Re: (Score:2)
Re: (Score:2)
Re: (Score:2)
So your view is that human infants don't have intelligence?
It doesn't matter what people might call a computer that acts like a baby, the point is whether intelligence is present or not, and if you believe intelligence isn't something that present in the human animal from birth, then sure, say equally dumb computers don't have intelligence, but most people believe that intelligence is an inherent trait in the human animal. Whether it grows from a low point doesn't really matter - the point is that intellige
Re: (Score:2)
As long as you live in the fantasy that human brains are magical and computers are not there's no point in having an argument, because you've defined the answer "Humans are intelligent and computers are not, hence anything done by computers is not proof of intelligence." rather than the question: "What is intelligence?" and failed to make any measurable definition or criteria. It's like saying humans have souls and rocks don't and expect the debate to be anything other than theology and philosophy.
The thing is, we've already defined these things, you just haven't looked it up or accepted it.
The kind of AI that wins games is called Applied AI (or Weak AI) this is very good at working within parameters, applying rules that do not change to data. Where it strugles is when rules aren't clearly defined or change. If someone were to cheat at the game the AI was playing, the AI would likely be unable to handle it. This is why games like Chess and Go are good tests for weak AI, the full parameters of the
Re: Out of date article^W summary (Score:2)
Re: (Score:3)
No, a bunch of people sat down and wrote software that lets the computer _learn_ to play Go. AlphaGo doesn't have a big list of "if board state X then do Y." That's precisely why Go was an interesting problem for AI to tackle -- the rules are completely deterministic to be sure, but the set of board states is so astronomical that you literally can't (and never will be able to) build a computer capable of recording them all.
Chess has been the go-to for "AI" for a long time because of its large state space.
Re: (Score:1)
Also, human brains that are not taught fairly quickly become incapable of learning.
Think of feral children- some who can't even learn to speak a language. Human brains need to be taught by other humans or their reasoning and logic and abstract thinking won't be there.
Re: (Score:1)
Well computers aren't capable of learning at all. They are mechanisms that react a specific way to instructions with no capability of independent or expanding thought. They don't gain wisdom and experience. They can't contemplate. They aren't conscious. They aren't alive. They aren't intelligent.
Re: (Score:1)
Computers are capable of learning now.
They passed being mechanisms that react in specific ways about 3 years ago.
They are not yet capable of independent thought but they have discovered new ideas which humans had failed to discover in 3,000 years of investigation.
They don't gain wisdom.
They do gain experience about particular subject matters.
They dont' contemplate.. the same way.
Most humans driving down the road are not conscious and are not intelligent. A small driving routine is executing and the rest of
Re: (Score:3, Insightful)
All of these articles are ridiculous. AI does not exist and won't exist for a very long time, if ever.
Re: (Score:3)
If you define AI as a computer sentience, it will only ever exist in the world of sci-fi. I prefer more relaxed definitions for what constitutes AI. If "AI" descibes an unattainable state, it ceases to be a useful term.
Re: (Score:3)
The issue is that the modern definition of AI is just far too relaxed. These days almost any computer system is called "AI" by the media, and often by it's developers. I pretty much expect the standard first program that everyone rights "hello world" to be classed as AI these days.
If people want to be taken seriously when they talk about AI, there needs to be some form of definition, and it needs to be more than just "this computer does something because it's programmed to do it"
Re: (Score:2)
I pretty much expect the standard first program that everyone rights "hello world" to be classed as AI these days.
"Hello world!" isn't good enough. To be AI, your program has to output, "I think therefore I am."
Re: (Score:2)
There is a form of definition: An AI is based on learning algorithms, rather than being directly programmed.
Of course there are still people who play fast and loose with the term anyway (especially in games where pretty much any non-triggered NPC behaviour is called "AI") but under the modern usage by people like Alpha and other major AI developers, what they mean is (usually) a neural network and other such learning algorithms.
Of course NN's aren't anywhere close to "true" intelligence -- even if they hav
Re: Out of date article^W summary (Score:1)
Re: (Score:1)
That was not, is not and will not be AI. Machine learning is not even close to actual AI. Just because a machine can store every possible combination of moves and appears to be "smart" does not make it smart nor intelligent.
AI will only happen once computers have done so much machine learning that they are truly making standalone decisions and not a decision based off of some rule set. Like voting for Trump, because and just because and no real reason or logic behind that decision. Once they make a decis
Re: (Score:3)
Let me preface this by saying that I agree with you.
That said, you undermine your argument here:
Until then, it's just a computer program doing a lookup against a huge data set and making a best guess as to the answer/choice
If a computer can truly "guess" than I'd say it really is AI. What computers do now is calculate probabilities and choose the most likely best move based on the algorithm in their programming.
And that's the crux of the whole thing. As long as a computer is doing exactly what it's programming tells it do do ("if X happens, do Y") then it's not AI. To be a true AI, it would have to come up with "X just happened, my
Re: (Score:2)
But AI are not lazy and will work to the death and not take breaks.
Normally if you hire someone Who is relatively stupid, they will be a problem because they are lazy too.
Re: (Score:2)
You run this program, you see a CPU Core peaking. The computer is putting all its effort into doing something that you wouldn't even notice after the screen gets full.
Seriously? (Score:2, Insightful)
That's because there is no AI. We got fancy algorithms that appear smart when guided by people, nothing more. They don't "think" and they are neither smart or stupid. That would require intelligence, which is missing from this whole equation...
"content moderators" (Score:2)
That has to be the most soul killing job in existence.
So are people... (Score:5, Insightful)
Look up any documented case of feral humans, either in the wild or confinement. If they have a few years first with parents beforehand, they tend to be OK after a period of catching up - but left completely "unprogrammed", they tend to be completely unable to cope.
Humans need interactions on several levels to "become" humans as we recognize them.
It's not at all surprising that computers would need some of that same kinds of interactions to be able to speak to us on our terms. We take a LOT of faulty shortcuts to real logic in order to play our roles in society, conversations, and our shared understanding of the world.
You can get a lot of that odd 'logic' just by building associations - but it takes a LOT of misunderstanding and correction before you can know if those corrections really work the way others understand them.
Ryan Fenton
Re: (Score:2)
AI don’t have instinct that people have. Feral people will still walk on two legs, they will come up with some forms of commication, and form some sort of social code.
There are some aspects that make us uniquely human that AI don’t have.
Re: (Score:2)
The instinctual traits you've mentioned exist in all (well most) living animals. So perhaps to improve the AI is to start with a foundation of code that can be considered the instinct. Then applied the same nurtural teaching that humans receive a they grow.
Re: (Score:2)
It's not at all surprising that computers would need some of that same kinds of interactions to be able to speak to us on our terms.
In an attempt to grasp why we should't anthropomorphize technology, we anthropomorphize technology.
Explaining machine logic to 'soft science' majors: futile
Re: (Score:2)
Re: (Score:2)
Look up any documented case of feral humans, either in the wild or confinement. If they have a few years first with parents beforehand, they tend to be OK after a period of catching up - but left completely "unprogrammed", they tend to be completely unable to cope.
Humans need interactions on several levels to "become" humans as we recognize them.
The big difference is that a human will learn of its own accord. Of course we guide it to get the kind of human we want in society but that is more a preference than an innate human ability. An AI at our current level of development will not learn or even act on its own accord, it is entirely dependent on a human to give it instructions and more importantly, give it parameters on how to carry out those instructions.
The fact AI do not know how to handle an undefined scenario without human input is why I f
Re: (Score:2)
Most humans do need guidelines, and parents to enforce them in their early years, police a bit later on.
No kidding (Score:5, Insightful)
Re: (Score:2)
Some of my favorite wtf *real sentences* that I've heard people say with a straight face in just the last week about AI:
-"This is a neural network, it works just like the human brain, with neurons, dendrites, and axons!"
-"The neural network evolves, we have species and genomes, just like the process that produced human intelligence!"
Problem of course being there is legitimate good work being done in the field, that will likely get flushed with the BS when the hype curve lets the industry down when folks rea
Re:No kidding (Score:5, Funny)
News Flash! Recently Discussed "Artificial Intelligence" Unmasked as Stock-Manipulating Hype
Yesterday, insiders from numerous think tanks and advanced computing companies came together to announce to the world that everything they've heard about Artificial Intelligence (or "AI") over the last few years has been false. "There are still no computers that can think unique thoughts on their own. It's all techno mumbo-jumbo and marketing speak to convince investors to invest in one company or the next. In fact, you may have been part of the entire effort to make AI seem more real.," said the spokesman for the group, Nerdy McSoontobejobless. "Chances are that you're in on the act, but just don't know it. If you've ever been asked to prove that you're 'not a robot' by selecting squares that include street signs, you're basically spoon-feeding an database algorithm what a sign is so that standard text-recognition software can figure out what the sign says."
"Oh ya," another representative amended, "Fully autonomous vehicles are still nowhere near ready for mass adoption. It's still going to be a decade or more until they're ready for personal ownership and, when they are, they're going to be extremely expensive."
The NASDAQ has dropped 15% since the announcement.
Re: (Score:2)
Of course, some implementations work best with humans in the loop. Doesn't mean every piece of AI software will need it.
Actually that's exactly the point. EVERY piece of AI software DOES need it. In this case "humans in the loop" doesn't mean interacting with the program in real time, but it does mean that every single possible thing the computer can do must be programmed in some fashion. Computers don't "learn" either. They load a data set that was given to them in some way, and then they use their programmed algorithms to interpret that data set and perform an action based on it. It's that algorithm that is the "Intelligen
Re: (Score:1)
No.
Alphago and Alphago Zero are learning from experience of playing the game.
It's learning by definition. It's not a program. It's not doing loockups.
It's not programmed. The base conditions were set but other than some instrumentations, the humans don't know why Alphago does what it does because they haven't played 134,000 games of go since breakfast.
Alphago Zero surpassed every human grand master go player in less than 100 hours. Human go masters are now learning new plays from it and Alphago.
Re: (Score:2)
No, it's not "learning" anything. It's a lookup table, it was programmed to run all the possible moves so that it can calculate which moves have the highest probability of success in every situation. It can beat a human because no human can remember that many possibilities and calculate them on the fly, but given enough time a human could follow the exact same algorithm and get the exact same results.
If you take that same program, without any modification, and tell it to play chess instead of go, it won't h
Re: (Score:2)
They taught it how to play go by letting it watch games being played. The second version learned by playing against itself.
There are video game playing AIs that learn to play pac-man or other classics by trying things and seeing the results it makes on the screen. They don't even define where the score is, or even that there is a score. They let the algorithm figure out what makes for a better result and it learns to play the game with no rules given to it. If you give the same algorithm a new game it can l
Re: (Score:2)
They taught it how to play go by letting it watch games being played. The second version learned by playing against itself.
So in other words, they inputted a lookup table.
There are video game playing AIs that learn to play pac-man or other classics by trying things and seeing the results it makes on the screen. They don't even define where the score is, or even that there is a score. They let the algorithm figure out what makes for a better result and it learns to play the game with no rules given to it. If you give the same algorithm a new game it can learn that one just as easily.
Except that's not at all how it works. You always are working from a data table. Either the table is from previous game play, in which case it will be incapable of performing any move that it has never seen done before, or the table is from playing against itself, in which case it needs to know the rules, and then will try to develop every possible combination of legal moves. Either way, it simply calculates from the table what move has the highest success rat
Re: (Score:2)
I'm not sure what you are calling a lookup table. Having an AI algorithm try things and remember what worked and what did not work does not sound like any kind of input table. The algorithm learns how to play by trying things out. I don't consider it intelligent, so I would suggest you stop confusing things by talking about unrelated things, this conversation is about learning, not about intelligence.
I remember one of the mario playing bots figured out some sort of cheat jump that it did not watch anyone pl
Re: (Score:1)
Your post shows you are completely ignorant of recent advances in A.I. research.
You need to go back to school.
Try the following.. read up on alphago and then go to youtube and watch Go masters observe and comment on Alphago's play.
Then read up on alphago zero.
When you can comment more intelligently than a chatbot... come back and make a more informed criticism.
There are significant limitations on A.I. but it's moving fast. It may never become conscious. Or we may develop an A.I. with emergent consciousne
Re: (Score:2)
I've looked in to it.
Alphago impresses people because it manages to make moves they hadn't thought of. It can make those moves because it has a lookup table that it has built of as many possible moves as it could, and can calculate the likelihood of success of each move based on it's algorithm. It is no more intelligent than the paperweight sitting on my desk. You can't take Alphago and ask it to perform a completely different task without re-writing the underlying code because it's not actually capable of
"disastrous outcome?" (Score:1, Offtopic)
It has become increasingly apparent over the past year that building systems without humans "in the loop" -- especially in the case of Facebook and the ads it linked to 470 "inauthentic" Russian-backed accounts -- can lead to disastrous outcomes
Because ... a few tens of thousands of dollars worth of FB ads actually caused people who loved Hillary to suddenly vote for the person they hated? This non-sly bit of editorializing (that the Clintons not regaining power to sell access from the White House was a "disaster" brought on by social media externalities) is laughable. Insufficient human involvement in FB's ad-processing may indeed have allowed some pot-stirring foreigners to run ads, but they also allowed an endless stream of domestically-passed
Re: (Score:1)
Re: (Score:1)
Trump's victory is the disaster.
I'm guessing you made big bets on the stock market going down. Or you make a lot of money based on millions of Americans having to wade through an incomprehensibly complex tax code. Or you employ a lot of manual labor that comes illegally across the border. Or maybe you make a lot of money abusing the H1B visa system? Be careful how you gamble on the next couple of election cycles.
Re: (Score:1)
Re: (Score:1)
Really? (Score:2)
That's just what they want us to think.
Actual Intelligence is similarly useless (Score:2)
You're kidding right? (Score:1)
It will impact their profits?!? C'mon...Facebook doesn't deserve the profitability it currently has. I've said it before and I'll say it again, Newspapers wouldn't sell without the important news and ideas that journalists provide to be printed on their pages; nor would those newspaper companies make any money from advertisements. Guess what; journalists / content creators get paid.
Not so on Facebook. The people who post their content on Facebook, or the articles that we link to aren't being properly co
Without human training, HUMAN intelligence is dumb (Score:1)
Feral children are what results from a human brain without human training.
Humans are born with a very plastic and adaptable brain.
WIth human training- it can be come a brain surgeon or theoretical mathematician.
Without human training- it eventually can't even learn to speak and is largely incapable of higher reasoning.
A.I. can't teach generally yet but properly configured once- it won't need humans again.
Just like humans.