WSJ Overstates the Case Of the Testy A.I.

mbeckman writes: According to a WSJ article titled "Artificial Intelligence machine gets testy with programmer," a Google computer program using a database of movie scripts supposedly "lashed out" at a human researcher who was repeatedly asking it to explain morality. After several apparent attempts to politely fend off the researcher, the AI ends the conversation with "I'm not in the mood for a philosophical debate." This, says the WSJ, illustrates how Google scientists are "teaching computers to mimic some of the ways a human brain works."

As any AI researcher can tell you, this is utter nonsense. Humans have no idea how the human, or any other brain, works, so we can hardly teach a machine how brains work. At best, Google is programming (not teaching) a computer to mimic the conversation of humans under highly constrained circumstances. And the methods used have nothing to do with true cognition.

AI hype to the public has gotten progressively more strident in recent years, misleading lay people into believing researchers are much further along than they really are — by orders of magnitude. I'd love to see legitimate A.I. researchers condemn this kind of hucksterism.

Comments

"No idea how... the brain works"

[Improv]

I'm calling the poster here out as being full of shit. As someone who's done neuroscience research, the idea that "Humans have no idea how the human, or any other brain, works" is bollocks. We have a reasonably good idea on the large scale, and in certain areas (such as the visual cortex), that understanding is quite far along. There are frontiers to our knowledge, but human understanding of brains is well on its way. Poster needs to pick up some neuroscience textbooks and get clued.

As a particular recommendation, I'd suggest Kolb and Whishaw's "Fundamentals of Human Neuropsychology"; it's an excellent textbook.

Re:"No idea how... the brain works"

[Improv]

Probably not - weak AI is typified by directly encoding domain knowledge on human capabilities into state machines, not typically meant to be neuroplausible or human-like. I believe the substrate here is wrong - real organisms learn (either as individuals or through generational building/encoding/selection towards instinct) how to do these things, and that knowledge is integrated. I don't think it'd be easy or likely that weak AI research methods will produce an integrated being with all these capabilities.

I'm sticking my neck out a bit here though; I'm not sure that weak AI research would be useless. Sufficiency versus usefulness is a complicated topic.

Also, my research was in neuroscience (led by cognitive modeling), not AI. It's a neighbouring field, but take what I have to say with at least a grain of salt.

Re:"No idea how... the brain works"

[Improv]

I agree that we don't have the full picture. That's not what mbeckman was claiming though, and saying "we know very little" because we don't have a particular achievement is an unjustified conclusion.

Re:There are ideas. Here's one.

[mbeckman]

Fyngyrz, by "some ideas" you mean "some theory". And in your case it's a theory with no research and no testable hypothesis. When I say "no idea" I mean literally we have no demonstrable understanding of any one single cognitive function of the brain. Any brain. We don't know how a gnat processes tactile information from its antennae, how a fly integrates spacial information while flying, or even how a planaria stores its memory of a maze. Human brains? We've got nothing.

Re:Teach vs Learn

[goose-incarnated]

Yes it does matter. If a piece of software does what it is programmed to do, in the direct sense, then it is not AI. If it can learn to respond or act in a manner that is not directly programed to do, then you are seeing whiffs of AI.

Using these goalposts even real intelligence, nevermind AI, would never meet the standard - if it has been directly programmed to learn new responses, ilke humans for example, then you would still fail it as intelligence using this criteria.

How about if what you directly programmed it to do was to write code to handle unexpected situations/inputs/etc? Perhaps in an iterative fashion, using previously gathered data? Using code fragments that are reassembled in new combinations, testing each mutation for success against the inputs? Because AIUI this is what the majority of chatbots *currently* do - use previously acquired data to refine their outputs.