Hacker News Reader

Comment by harshreality | Hacker News Reader

Reads like an article written 5-10 years ago, not to mention it doesn't mention neural networks (by name) even once. Forget the singularity... the article doesn't mention the singularity except in the title. This is a "computers aren't special they can never be like us" article, based on an interview with a single person who seems more focused on robotics than neural network research.

It doesn't prove anything about the future of AI to give anecdotal examples of computers failing at being intelligent (nobody has claimed that they are human-level at most things, yet), and then draw conclusions from that biased data.

It's more like a half-interview half-essay of a discussion with an AI skeptic who doesn't bring up any of the modern (as of 2015 when the article was published) approaches to AI, and goes on irrelevant tangents like the uncanny valley (which has to do with HCI, not the singularity or AI creativity, which are the subjects of the article).

View on HN

Every article claiming we're on the cusp of the singularity reads identically to those written 10-15 years ago too.

This reads this way because computers haven't made large strides in any creative endeavor in the last 10-15 years. Why should that be a point against the argument?

The majority of the discussion on the topic is not claiming we are on the cusp. It is claiming that it is a) a danger; b) likely to be fast if it happens.

Almost maximally irrelevant to a claim that nothing significant has happened recently, so not a problem.

I'm not sure that this argument is entirely sound, given how incredibly slowly humans move as a species to cope with anything new, and how quickly this particular kind of "new" would be likely to emerge. I don't know about you, but a few generations of ongoing discussion seems downright necessary to me, if we have a hope of not making some catastrophic mistakes.

Computers can now take images and produce output in the style of major artists eg. http://www.telegraph.co.uk/technology/picture-galleries/1184...

You may say that's just an algorithm but it's way better than I can do personally. That was all figured out within the last decade.

Also the computer art was produced by artificial neural networks, my attempts if I made some would be done by regular neural networks. Who's to say the computer stuff doesn't count as creative but mine does?

Computers are way better than me at a lot of things: arithmetic, remembering things, surviving on a diet of plain electricity, just to mention a few...

Transferring local patterns between images and playing Go have recently been shown to be a few more of those things. But how far will that generalize? Nobody knows.

Computers cannot take images and produce output in the style of major artists. If you think they can, give them realistic art and watch them embed eyes from the source image into your mouth.

This is not to say that it can't ever be done, just that right now it isn't done by people claiming they can do it. That articles with results similar to 100x cheaper random texture synthesis from 15 years ago get published and hyped in the general press tells you how much of an AI bubble we live in right now.

Yup. The linked image gallery is actually a pretty good argument against computers being able to take images and produce output in the style of major artists. If you asked a human to paint the Golden Gate bridge in the style of Matisse's "Woman with a hat", they probably wouldn't interpret the request as "make part of the sky look like a hat"[1]. If you asked a human to draw a group of buildings by a canal in the style of JMW Turner, they wouldn't make the buildings look like abstract waveforms[2], even if they were only familiar with his stormy seascapes and had never seen how he actually drew groups of buildings by a canal[3]

[1]http://www.telegraph.co.uk/technology/picture-galleries/1184... [2]http://www.telegraph.co.uk/technology/picture-galleries/1184... [3]https://www.google.co.uk/search?q=jmw+turner+venice&source=l...

heh a decent point, but to be fair the task really isn't "make a picture in the style of <artist>" but "make a picture in the style of this one other picture". I'm not particularly convinced an otherwise naive human could separate style from motif in such a limited setting.

It mainly demonstrates that the network learns rather abstract characteristics of the original picture, to be possible to factor them out. Hardly a way to create new masterpieces, but a nice demo of computer vision advancement.

Style transfer is just a complicated method for tracing one picture onto another.

what? eyes?? really? aren't you confusing the deep dream work with the style transfer work? I remember the dreams thing having eyes everywhere.

Can you show me an example? I'm thinking of something like this: https://github.com/jcjohnson/neural-style though I think there was a webapp to upload your own and try it out that I couldn't find now...

In the last five years, AI/ML has gotten better at some things it was previously mediocre/good at, but what progress towards creativity or singularity has there been? None. DL is not an enabling technology for creativity.

This is something it amazes me more people don't see. I've read a lot of the scientific literature on human creativity (there isn't much) and it is quite clear that we don't understand it at all.

Additionally, there is a transformation that takes place in our brains from information to knowledge. Our understanding of this process is also pretty weak.

Both of these problems are fundamental to the creation of true, autonomous AI.

Both of those problems might be something we need to solve, or it might emerge out of the complexity of a sufficiently advanced AI system. If the latter, we might not seeing coming for exactly the reasons you've outlined as making that emergence unlikely.

> Both of these problems are fundamental to the creation of true, autonomous AI.

What makes you say that? I could see your statement being self-evidently true only if you meant "fundamental to the creation of an anthropomorphized AI." Of course, even that glosses over the fact that there are already strong indications within the AI community that the first general purpose AI will likely not "think" even remotely in the same fashion as humans do.

> I've read a lot of the scientific literature on human creativity (there isn't much) and it is quite clear that we don't understand it at all.

I don't agree that this is true in any way that precludes the creation of "true, autonomous AI". We may not understand the specific mechanisms of how our brains give rise to creativity and the resulting "implementation artifacts", but the fundamental part of creativity (and I'd argue, the part that's relevant to AI) is the ability to efficiently search huge but structured hypothesis spaces and inductively learn new structure (very likely with some stochastic element involved).

That is something that we not only have an understanding of, but have implemented, more than once. By way of examples:

http://www.popsci.com/scitech/article/2006-04/john-koza-has-... http://aliciapatterson.org/stories/eurisko-computer-mind-its... https://psmag.com/triumph-of-the-cyborg-composer-620e5aead47...

We can create true, autonomous Intelligence, even though we don't understand how it really works. I predict that - we'll be able to create AI and train it just like we do with a human being.

Does playing Go and beating world champions count? That seemed a creative endeavor to me.

I would say those are analytical, but not creative.

As soon as a computer becomes better than the human (either best or average one) in something which was considered as a sign of intelligence, immediately texts will be written why this is not a sign of intelligence and why intelligence is something completely different. This repeats many times. The same holds for creativity.

I have rather the impression that lots of people who write this kind of papers rather want some computer program who is just like them - just a little bit more clever or creative (call it narcissism) instead of some computer program that is "intelligent" or "creative". These people don't understand that one shouldn't judge a fish by its inability to climb a tree.

That may be, but it may also be that when computers become better than a human at something, that happened because we gained a greater understanding of how to make that happen, and that greater understanding is what leads us to stop considering it a sign of intelligence.

Hero of Alexandria, who lived from 10AD to 70AD, is famous for his mechanical inventions. One that I recall was a mechanism that opened the doors to a temple when a fire was lit in a brazier outside. To the people of the time, who didn't understand mechanics, this was a sign of the gods and the everyday miracles they would perform. But to anyone who knew how it worked, it was a simple mechanism.

A computer that plays a complex game, or creates art, or understands spoken queries and answers verbally, or identifies things in photos, or drives a car, seems like magic to most people. But to software developers who understand the algorithms behind those actions there's no magic, and no intelligence. It's just software.

The promise, and claim, of AI boils down to this: that what's going on inside of our own heads is also just software. Maybe that's true, maybe it isn't. Even if it's true, there's no guarantee that equivalent software can be run on the artificial hardware we've produced. So yes, we keep raising the bar, but I think that's because we keep discovering that the bar was actually already higher than we thought it was.

Reference the AI effect:

https://en.m.wikipedia.org/wiki/AI_effect

well if intelligence is recognising and extrapolating from regular patterns in your sensorium, surely the AlphaGo system is rather intelligent; even just playing against the search-less pattern subsystems on the version that beat Fan Hui (way weaker than the one we saw in the Lee Sedol match by Google's internal ELO assesment) almost any non-professionals would get crushed. So it clearly learned a great deal from both go game records, and from experience of playing.

But isn't a question of creativity rather different from the matter of intelligence? I'd love to see recommendations on papers that try to formalize creative AI systems -- but I'd expect something with less focus on exploitation and more on exploration than a goal-directed game engine. Maybe using some novelty signals or the like.

The fact that people redefine what AI whenever reached is is a valid point; but that does not mean one should lump all sistems into the same basket.

Something like a pop music composing and light novel/screenplay writing bot at least seems like a doable goal, yes, but it def isn't AlphaGo.

Move 37 game 2 by AlphaGo against Sedol:

>But at the same time, those of us who watched the match inside Seoul’s Four Seasons hotel could feel the beauty of that one move, especially after talking to the infectiously philosophical Fan Hui. “So beautiful,” he kept saying. “So beautiful.”

(Fan Gui is the three-time European Go champion)

If any human played that move it would be described as very creative. https://www.wired.com/2016/03/two-moves-alphago-lee-sedol-re...

That's a human emotional response to an observation. We do it all of the time; we see beauty in landscapes, grace in the movements of animals, peace and tranquility in the sounds of a babbling brook. None of these are really out there in the world; they're reflections of our own minds upon nature.

A human player might have made that move in a burst of creative inspiration. Or they might have made it after careful analytical deliberation. Or it could have been random desperation. If they don't say why they did it, observers are free to interpret it any way they like. That has no bearing on what actually happened.

hm, used to be said for chess too; a uniquely human creative skill.. I love the Go algorithm (there's 3 basic ones in there really), but no aspect of it seems creative.

To try to see it in a grossly simplified step by step fashion: in the abstract, does seem similar to how a human would think of the game though; consider promising moves, where promising is some fuzzy intuition based on the pattern of the board -> for each of those, read how the responses might look like (which the computer does way way more than the human), and what you discover using this suggests other promising moves -> and at some depth of each such read, evaluate the board position; who looks ahead, which too is some fuzzy-ish intuition based on the board patern (though there are various heuristics humans would use when learning this). Computer also evaluates further, a number of fast playouts to the end and mixes this with that pattern-based evaluation.

Is this the kind of thought process you'd associate with creativity? Seems search-based, analytic and strategic; only funny thing is how evaluations are pattern-based nebulous things, but that seems an implementation detail. In some situations, you may come up with evaluation procedures like say material advantage in chess and code the relevant decision procedure, in others, you'd "code" the decision procedure implicitly, by specifying a dataset that exemplifies the desired function, and use statistics to come up with that function. Increases the kinds of decisions you can implement tremendously as you can leverage implicit rather than just explicit knowledge, yes -- but conceptually kinda the same.

Human creativity is not some magical thing. Plenty of humans create things of absolutely no value to their surroundings while being creative.

Creativity is much more likely a human way of understanding the ability to create new new things with some sort of a purpose or intent.

But humans have limits to their ability to create things too. There are things that only machines can create and perform.

They can create things only in the sense that they receive a set of inputs and produce deterministic outputs according to a process explicitly specified by an prespecified algorithm. It's not "creative" in the conventional sense of the word, but perhaps humans do essentially the same thing in a much more advanced way.

Yeah I guess the point is that human creativity is expressed in some sort of skills/trade which are learnable whether it's playing an instrument, painting or writing a poem.

The state of human creativity is the accumulation of hundreds of thousands of years of trial and error.

It can be put into an evolutionary context and so in my belief, is accessible to computers who now can also learn.

Go back to 1950, you will still find arguments that we are a few years away from general purpose AI.

The arguments have not really changed.

Depends who's arguments you look at. The Kurzweil arguments have always been turing test 2029, singularity 2045.

So say general purpose AI about 2030. So it was 80 years in the future in 1950, 14 years now and so on for other dates. It depends a bit on the definitions you use but around then.

I prefer not listening to techno astrologers, and Kurweil is one. By and large the more one contributes to AI, the less one is an techno astrologer.

Kurweil's contribution to AI pales in comparison to the likes of Norvig etc.

Though sometimes nicely written, nautil.us articles basically always disappoint in terms of actual substance. Not sure why they're on the frontpage of HN so often.

The presentation of the page and the aesthetics has a lot to do with it.

That same statement is a common complaint about HN.

Actually, he does talk about modern approaches to AI. He just says, they don't change shit. You know when neural networks were invented? In 1943.

Reply on news.ycombinator.com