Googling the drawbacks to progress in AI research pops up points like:

  1. Machine learning requires vast amount of data
  2. We don't really even understand how neural networks work, we just know they do.
  3. Neural networks can be programmed to do only one thing at a time.

Is it possible that today's AI designs could actually be better than we imagined if it ran on fast enough processors and had enough data to process?

(I'm referring to neural networks because that's what I've read up on, but other AI designs are also included in this question.)

Is it really true that neural networks need to be reprogrammed to learn new tasks, or is that just how we are training them. Imagine a child spending his entire life playing nothing but chess, and eating and drinking. No conversation, no education, nothing but chess. Eventually we may end up with a highly skilled chess player, but then the child may find it difficult to learn the simplest of new tasks.

Maybe a neural network needs to be taught how to do 20 different things before it can even realise that all the tasks it is being given are not the same, and hence it must learn differently in different tasks.

Is today's AI unable to multitask because it isn't complex enough at the design level, or simply because we lack the millions of hours it would take to teach it something meaningful?


closed as off-topic by L.Dutch, StephenG, Vincent, Paul TIKI, Vylix Nov 18 '17 at 17:54

This question appears to be off-topic. The users who voted to close gave this specific reason:

  • "This question does not appear to be about worldbuilding, within the scope defined in the help center." – L.Dutch, StephenG, Vincent, Paul TIKI, Vylix
If this question can be reworded to fit the rules in the help center, please edit the question.

  • 1
    $\begingroup$ This sounds more like a computer science question rather then worldbuilding. P.s. I don't get what you mean with point 2. $\endgroup$ – L.Dutch Nov 18 '17 at 15:17
  • $\begingroup$ Downvoted: * Don't see the relevance to world building. * Primarily opinion based. Suggest rephrasing the question as "How can a given AI system be shown incapable of sentience" and ask it on computing science or philosophy. $\endgroup$ – Sherwood Botsford Nov 18 '17 at 15:22
  • $\begingroup$ Can you define what "appearing sentient" means to you? It's a very tricky and very loaded phrasing. $\endgroup$ – Cort Ammon Nov 18 '17 at 15:27
  • $\begingroup$ Wow, another downvoted question. I know, I know. $\endgroup$ – ghosts_in_the_code Nov 18 '17 at 15:33
  • $\begingroup$ @CortAmmon Okay, I've deleted the word sentient. $\endgroup$ – ghosts_in_the_code Nov 18 '17 at 15:34

(This started as a comment, and evolved into a lengthy explanation. TLDR; AI is not what the OP thinks it is, and the question shows strange misconceptions about what ANNs are, how they work and what they can do.)

What is Artificial Intelligence?

Artificial Intelligence is a sub-discipline of computer science; by and large, artificial intelligence plays within computer science the role that philosophy plays within science: that is to say, if we know how to approach a class of problems, then that class of problems gets its own sub-discipline (such as numerical calculus, linear programming, or sorting and searching) and if we don't know then the class of problems is shunted to the wastebasket sub-discipline of "artificial intelligence".

As time passes and we discover ways to approach various classes of problems they move out of artificial intelligence to acquire sub-disciplines of their own; for example, thirty or so years ago chess-playing and grammar checking were considered to belong to artificial intelligence; but now there are efficient ways of doing both, so they are no longer considered to belong there. We see this happening before our eyes with image classification and face recognition; ten years ago both those classes of problems would have been considered to belong to artificial intelligence by any worker in computer science, but today we are very close to having excellent efficient methods to classify images and to discriminate between human faces so that today most practitioners would consider them to be distinct sub-disciplines. As Wikipedia says (quoting Douglas Hofstadter, the author of the famous Gödel, Escher, Bach), "AI is whatever hasn't been done yet".

Neural Networks

I have no idea why the OP believes that "we don't really even understand how neural networks work". This is patently false. We know perfectly well how artificial neural networks (ANNs) work; that's why we can use them successfully. It may be the case that the OP is referring to the inability of a neural network to provide explanations for its results; this is indeed a big drawback, and it's inherent in how they work; and we definitely know what they can do and what they cannot.

ANNs are good for classifying stuff, where the "stuff" can be anything that can be represented in a computer. They can be trained to discriminate between pictures of sheep and pictures of goats, or between phonemes, or between grammatically correct and incorrect sentences. That is essentially all they can do, and you need a new set of parameters for each and every application. The big drawback is that the result comes "as is", with no explanation, unless one is ready to accept a large set of numbers (representing the weight coefficients assigned to the myriad inputs by each node) as an explanation; and when they fail and misclassify the stuff, they fail spectacularly: Ars Technica had a delightful article on how self-driving cars (which use ANNs to recognize traffic signs) can be confused by stickers pasted on stop signs (quoting a serious study by Ivan Evtimov, Kevin Eykholt et al.).


The OP says that "neural networks can be programmed to do only one thing at a time", which is trivially true. But the OP forgets that an ANN is a mathematical structure implemented in dedicated hardware as a large matrix of weight coefficients; you can have as many active (loaded) ANNs as you want working in parallel, and you can have many more stored on disk ready to be loaded unto the hardware. This is not a serious limitation.


Not the answer you're looking for? Browse other questions tagged or ask your own question.