AIs getting confused

Well, do you ever watch YouTube videos that you’ve seen before? I know I do, for instance because I often let music videos from YouTube run in the background, and I have some favourites there. So the algorithms have, presumably, learnt about me that having watched a video is not a reason for not them not to recommend it to me. I agree with you that it would be a sign of bad performance on the algorithms’ part if they recommended you a video that you’ve seen before even though you never re-watch anything; so my guess is that, as a matter of fact, you do re-watch videos more frequently than you think you remember you do.

This does not fit with my own (non-technical) experience of machine learning - for example with text to image GANs, one can ask for (and sometimes receive) things like ‘A tortoise on toast’ and as long as the algorithm has been sufficiently trained with tortoises, toast, and ‘things ON things’, it will have a go at composing it. Not always successfully, but with enough success to make me believe that there is some generalisable learning.

I mean, for example, look at this:

I agree that AI is narrower than human intelligence; but it keeps getting wider. An example for that is the fusion of speech recognition and automatic translation. Both are things that have been worked on for a long time in AI, and algorithms to do either have been around for decades (though with poor results in the early years). Both are, in themselves, a narrow mimicry of one particular task humans can do. But take a reasonably well functioning speech recognition system and a reasonably well functioning translation system, combine them, and voilà, you have an app which, in real time, interprets spoken language from one language to another. Which is still relatively narrow, compared to the overall range of tasks humans can do, but it’s wider than either of the two before they were combined. In the next decades, we’ll see a lot of those mergers of previously separate algorithms to widen the range of human tasks that can be performed algorithmically.

If we were to distil it down though, is that not true of many things humans do? I’m able to speak
because I learned to mimic what my parents could already do. That mimicry is happening because of processes in my brain that are just processes.

I’ll just add that there’s nothing to stop Google from including a button that let’s you inform the AI that it did it wrong. If they did, the AI would be able to continue learning and improving.

The problem with that, though, is incidents like these:

When the AI is trainable by users, some users will take advantage of that and teach it to do the wrong thing.

Crowdsourcing the teaching could (in theory) be immensely quick and powerful, for creating near-perfect AIs. But, in the real world, the companies have to generate their own training data and that means hiring people to sit in a room and generate examples, then have another group go through and sift through that to remove anything problematic, and then have another group go through to make sure that there wasn’t any unconscious bias put in through sample selection, etc. This results in you having a fairly small set of training data and thus having an AI that can’t act correctly under all scenarios.

This reminds me of the news article from about 2002 “Help! My TIVO Thinks I’m Gay!”

Apparently by setting it to record Will and Grace his TIVO started suggesting and pre-recording dozens of gay-themed TV shows. There’s a follow-up article on how to do an intervention on your TIVO.

Occasionally, but not 10 minutes or even days later, more like months. I swear I’ll finish watching a video, go back to the homepage, and there it is in the recommendations. Useless. Now at some point the red bar disappears from videos you have watched before. I’m not sure how the algorithm handles that.

YouTube sure doesn’t seem to be listening to my feedback about this. I just got 20 videos in a row I had already seen, many of which in the last 24 hours. I report them as “Not interested > already seen it” because I was getting tired of so many of those, and now I get a lot more.

I’ve given feedback many times about how annoying this is, but YouTube seems to care as much about that as they do about their pornbot spam problem.

I largely disagree with this. Goalposts have mostly been moved by self-styled AI skeptics, like the late Hubert Dreyfus who famously declared that computers would never (regardless of computing power) be able to play anything better than a very elementary child-like beginner level of chess. He had to eat his words and move the goalposts when he accepted a challenge and was himself soundly beaten by one of the earliest chess programs to demonstrate real skill and apparent strategy.

I disagree with the idea that intelligence requires generalizable knowledge. This sounds like a skeptic’s fallacy arising from an unwillingness to accept that intelligence in a narrow domain should be regarded as “true” intelligence, even when it’s effectively solving real intellectual problems. It seems to me that it’s more reasonable to establish a priori that if some entity – whether human or machine – is capable of performing a certain task of reasoning that would be commonly regarded as requiring intelligence, then it should be deemed intelligent in that domain, whether human or machine. Otherwise the goalposts just keep moving forever and the whole argument becomes meaningless, simply a circular declaration that “if a machine does it, it’s not real intelligence”.

I disagree with this, too. The skeptics weren’t “underestimating the power of a focused, brute force approach” as that’s not the way successful AI systems generally work. Indeed, the chess challenge that beat Dreyfus occurred in 1967, and ran on a DEC PDP-6, a machine first released in 1964. This machine was absurdly slow by modern standards. No, what the skeptics were underestimating was the power of heuristics and clever algorithms to dramatically trim the search tree when evaluating chess moves. More broadly, they failed to understand the potential of stored-program computers to effectively create the behaviours of human intelligence. There’s no question that faster computers have been a great boon to AI (lots of computer power was involved in creating the IBM Watson Jeopardy champion) but that’s just one factor of many.

Sorry for the digression, but I felt I should respond to those points.

To my mind text-to-image is a less complicated task than image description, at least by the success criteria it seems you are applying. The extent to which this “learning” is generalizable also seems to be built into the task. It seems my use of generalizable in the context I meant it isn’t generalizable to others. :wink:

Yes, but you aren’t limited to that one particular task and the mimicry was only part of the learning process. I’ve taught math for over a decade and find it a complete waste of my and my students’ time if they never move beyond rote repetition of an algorithm.

And this sounds like an unwillingness to accept that “intelligence” is a word like any other with varying definitions. To me intelligence is close to consciousness in vagueness and complexity and I don’t think it valid to a priori establish tasks that could be completed that would show a system had one or the other. I think machine intelligence is much more likely than computer consciousness, but by how I think about the concept of intelligence, machines are nowhere near.

Text to image requires image description as part of the process - that’s how GANs work - one algorithm is trying to make an image of ‘tortoise on toast’, while a competing algorithm criticises it for accuracy.

One selected aspect of me is not the whole of me. Never said it was. Machine learning can perform some feats of learning that are similar to some of the things that humans can do. The mistake is to imagine that they can only do things they are explicitly programmed to do - the ability is in the emergent properties of the learning.

But a machine that has learned to paint pictures isn’t going to suddenly get dissatisfied and decide to teach itself piano - that sort of generalisation isn’t expected from anything like we’ve been discussing.

Decades ago I read an article in Analog about AI where it was comparing today’s attempts at AI with the early attempts at flight, the one where you see a flapping machine shaking itself to pieces instead of getting itself off of the ground. It wasn’t until, instead of trying to fly as birds do, the idea of a fixed wing and propeller was perfected that powered flight was possible.

He argued we need to come up with the intelligence equivalent of propeller-and-wing before we’ll get anywhere.

Depending on how long ago that was, that author might have been too pessimistic. The dominant techniques that have allowed machine learning to make its recent rapid progress are things like neural networks or statistical methods such as linear regression that have been around for a long time. In a way they’re inspired by the natural role mode of the human brain and apply similar principles, but they’re not a direct copy of it.

Yeah, it seems like if we make a machine do something that a human can do, it’s ‘merely mimicking’ and if we make it do something interestingly different it’s ‘not like a human’. Skeptics need to define in advance what are the criteria to be met, rather than this continual game of ‘yes… but’.

Edsger Dijsktra famously said that the question whether computers can think is about as irrelevant as the question whether submarines can swim. I don’t know exactly what he meant by that; but my interpretation is that one can have lengthy semantic discussions about whether the word “swim” should be applied to what submarines do, but that this does not alter the fact that submarines are very good at what they were built to do, and probably better in some respects than humans are in a comparable activity, no matter which word you apply to it.

I can’t help thinking if it’s whole-species loneliness. We want to see if we can communicate with animals; we want to try to communicate with extraterrestrials; we want to build things that can think like we do, so we can have someone to talk to…

I think the next AI frontier will be “it may be conscious, but how would we tell?” We will have built systems so good at imitating some attributes of human interaction that we will wonder for say, a Turing test, whether it’s simply complex analysis of context to provide responses and such, or whether there’s conscious action at work…

My friends and I used to joke that we knew plenty of humans who could not pass a Turing test.