AIs are now re-writing history | Hacker News Reader

AIs are now re-writing history

robertelliottsmith.com

116 pointsmooreds53 comments

The middle paragraphs are just further proof that talking to a blowhard like Cory Doctorow is a complete waste of time. The "all photos are fake at some level" slippery slope argument is beyond tiresome. The fact that something like white balance introduces a small amount of post-processing into photographs does not mean that all levels of photo manipulation are somehow equal. That's nonsense on its face and should be called out as such. You need a trampoline and a net to make that kind of jump in logic.

Exactly. If a dog views the same scene as me, and it presumably sees only black and white, did we see a different reality? No, we saw different colors.

Off topic, but just so you know -- dogs are colorblind (as compared to us) in that they have a more limited perception of color, but they don't just see black and white (and shades of gray)!

http://www.psychologytoday.com/blog/canine-corner/200810/can...

Well, it depends on his tone. When pondering a question, I feel like it's good to define boundaries and then work your way inwards. He sets the lower bound: all photos are modified in some way. We are okay with that. But we're not okay with this smile edit. At what point in between does it become not okay? It makes you think very specifically about why something makes you feel the way you do.

He might not be saying, "these two things are the same," but rather, "this situation is similar, what makes this one okay but not that one?"

I don't know if that's what he was doing though. But I find my discussions start off similarly. I reframe a question in the context of similar things that I already know how to feel about, and find the intersection points.

"I told him that algorithms are, without prompting from their human designers or the owners of the photos, creating human moments that never existed."

I'm not sure how anyone could make the argument that this occurred without prompting from their human designers.

This seems to be a continuation of a theme, where AI academics and theorists are so divorced from the realities of present-day engineering that they assume we're much closer to singularity than we are. I spend so much time making things display in the correct order.

"Smile!" is one config file away from putting his head on her body. It might be hilarious, but it's not AI.

Citation of hilarity http://manbabies.com/

I don't think its necessarily a question of how close we are to the singularity. I think it's a question of how close we are to the Orwellian Memory Hole.

But that was inspired by Stalin's photo manipulations, so it's not a question of "how close", we effectively already live in that world.

Stalin's oppression took quite a lot of man-hours.

This is really cool! I agree with the author a bit, however I think his concern stems from the fact that it was largely because he did not know it was going to happen.

I also think that Doctorow's dismissal of his concerns as "trivial on the scale" is silly and somewhat arrogant. Automated alteration of a picture to the point of making a new picture non-discernible from an original is a real different thing. In this case it was very subtle, but the application of this could really have some broad impacts. I recall for example the doctored photo from Iranian missile launches [1] or the Egyptian PM red carpet photoshop [2].

Yes this has been going on forever, but automating it makes the impacts really scale quickly.

[1]http://thelede.blogs.nytimes.com/2008/07/10/in-an-iranian-im...

[2]http://www.theguardian.com/world/2010/sep/16/mubarak-doctore...

How do you not know it will happen? I haven't used AutoAwesome since just after it was released, but at least back then it was an opt-in. Google didn't just automatically overwrite your pictures, you selected which ones you wanted to make AutoAwesome.

Has that changed?

I should note that I don't have any special information about AutoAwesome, so my knowledge is based on the blog post.

The author seemed to be surprised which makes me think it was unplanned, but I might have just been reading into it.

According to the Google help page referenced in the article, it is on by default, but you are notified when it happens. Pictures aren't overwritten; new ones are added.

It hasn't changed at all. I have it enabled, its funny sometimes. The 2013 year end video brought a tear to my eye, and it changed some beach photos by adding in snowflakes (must have thought the white sand was snow).

The title is a bit hyped, but the examination of implications of algorithms is interesting.

I appreciate the artistic license of saying the "AIs" are doing this, but this isn't an AI issue at all. This is a particular single purpose program "Smile!" which is producing results by combining pictures together to produce a picture the author never took with a camera.

He could have blamed Snidely Whiplash, or the greek god Zeus with just as much verity.

To generate this effect a program has to identify, that the photos are almost identical, that there are smiling people in them and then find a way to stitch the two photos together. I think that qualifies as AI.

It's AI in the sense that a pocket calculator from 1980 is AI. It doesn't know what it's doing, can only perform a limited range of calculations under tight constraints, has no awareness of whether or not the result is meaningful, and unless the authors went seriously overboard developing the suite, cannot learn from what it has done to improve the process.

It's pretty likely that the techniques being used fall under the term AI as used by people in the field. Actually, I'd be quite surprised if that wasn't the case.

Well, I am afraid that we are discussing semantics. The functions performed by the software, like facial recognition, are classical AI problems. Granted the software is certainly not any kind of general intelligence, and probably not even cutting edge AI, but it does still perform pattern recognition in unstructured data etc.

Ironically, parsing semantics is also a classic AI problem.

Twenty years ago what your phone is doing now was called AI. Now you call it "search engine" and "voice recognition". In the mind of people from outside the field, "AI" means "things computers can't do". The moment they are able to do something, it stops being called "AI" and becomes mundane.

It does, but in the article, the author is clearly using a grander sense of "AI" where computers are making autonomous decisions without our knowledge rather than just a clever heuristic. This is not that AI.

That's a very stupid AI, directed by a human to robotically go after certain categories of pictures and meld them.

https://en.wikipedia.org/wiki/Weak_AI

>Weak AI (also known as narrow AI) defines non-sentient computer intelligence, typically focused on a narrow task. The intelligence of weak AI is limited.

Typically people refer to algorithms that use machine learning, optimization, or other things once believed to be part of the field of AI (once we understand how to do them, they are no longer "AI".) People have been using the word "AI" like this for decades, to the point it's used far more to refer to weak AI than strong AI. When you talk about strong AI you generally have to specify it.

I would say that AI is very closely related to this though. Deep learning is taking over machine vision and simpler hand coded algorithms. It has the capability to generate good synthetic images, and it will probably be used for stuff like this a lot in the future, if it isn't already.

This an interesting example, but other photo processing techniques can really make the "commissar vanish" and jack up the creepy factor. The same photo software vendors that make your smartphone jpgs look uncannily good can remove moving objects from your vacation snaps. The person being chased by a cop across your Eiffel Tower snap will just vanish without a trace.

The shutter button is just a hint. The camera knows what you want to see before you do.

Doesn't this simply illustrate that as science advances, education of the general population must keep up to raise awareness of what's possible?

We've failed to educate people about many risks associated with the use of computers, but "photoshopped" images are a well-known phenomenon (though doing this to the extent described in the article, without anyone being aware of it, is a bit of a novelty).

And if all others accepted the lie which the Party imposed – if all records told the same tale – then the lie passed into history and became truth. - George Orwell

Some time ago I conducted a simple (albeit impossible in practice, for now at least) thought experiment. Let's say that average size of a JPEG image is 5 MB. Now, some machine could generate all possible valid JPEG files sized 5 MB. In some of those images there would be events which never happen, and never could (for example Usain Bolt running in 1896 Athens Olympic Games etc.) Of course, sheer number of such images is ginormous, but for me it's interesting (and scary) that something like this will maybe be possible...

A group of classmates and I spend about a week obsessed with this idea once. It's also cool to think that among all those images you would have e.g. the whole history of the world from all possible perspectives.

I doubt it will ever be possible, though. Even for tiny, grayscale images the amount of images you'd have to generate is absolutely impractical, not to mention the effort to distinguish noise from actual pictures.

Not to mention that the whole thing would be illegal. It'd have child porn, state secrets, libel, copyrighted software, etc.

Or a computer generate all bytestrings and run them as executables. One of them would post this comment -- or maybe it is!

> Of course, sheer number of such images is ginormous, but for me it's interesting (and scary) that something like this will maybe be possible...

It will never be possible. A 5 MB JPEG has 5,368,709,120 bits, which means there are 2^5,368,709,120 different possible images. In scientific notation, it's 1.302... × 10^1,616,142,483.

For comparison, there are about 10^80 atoms in the entire universe.

I suspect you can significantly cut down the space by limiting it to valid JPEG files.

You can cut it down but it still won't be feasible. Consider generating even just all possible 100x100 1-bit monochrome images. There will be 100 * 100 = 10000 bits per image, giving a total of 2^10000 possible images: about 1.995 * 10^3010 which is still way, way bigger than 10^80.

A 5MB JPEG image would have a much larger resolution and color pallete then this.

A 5MB JPEG corresponds to 12~16MP@24bpp, although many of the images would compress to the same result.

I'm not sure you realise the scales involved. 5MB (40Mb) is enough to create 1e12626113 different files, that's a number with 12.5 million digits

For reference, the number of atoms in the visible universe is ~1e80 (1e78 to 1e82), you'd generate 1e12626033 images per atom (doesn't look any different does it? Well it's got 80 zeroes chopped off). For a second reference, assuming you're a standard-size hooman you have 1e14 cells tops (estimate vary, "An estimation of the number of cells in the human body"[0] gave 3.7e13).

So no, it's not possible. Even if you remove a few order of magnitudes to account for invalid jpegs, let's say we cut it by a thousand orders of magnitude it's still 1e12000 images to generate[1].

They're also all contained in PI. And the thought experiment is but a variant of the infinite monkeys theorem[2]

[0] http://informahealthcare.com/doi/abs/10.3109/03014460.2013.8...

[1] if you can compute full tilt until the heath death of the universe you have ~1e47 seconds, you'd need to compute 1e11953 images per second… or 1e11873 images per second per atom in the universe

[2] http://en.wikipedia.org/wiki/Infinite_monkey_theorem

:-]

I thought of a similar thing for an mp3; a computer would produce every possible "song"..but even at an incredibly low resolution of 100 kilobytes for a 3 minute song, you're looking at 2^800,000 possibilities. (1e240000).

Another cool one is possible games of Go, ~1e768. This is one reason top computers still can't touch top humans at winning the game.

Big numbers are fun. :-]

I don't think the number of possible games has anything to do with the difficulty of programming computers to play Go. There are 10^120 possible games of chess, for instance - still 40 orders of magnitude larger than the number of protons in the observable universe - and yet computers roundly defeat humans at chess these days.

100% correct.

But what if we imagine/allow/invent the series of algorithms that 'pluck' a good picture of Usain Bolt from the set of 5MB files without inspecting too many (let's say O(log) or O(log log))?

That's only ~c25/~c5 bits of entropy to sort through, even less if the algorithm were to work in some sort of sparsity domain.

You remind me of the space pirate Pugg, as encountered in the sixth sally of Trurl and Klapaucius as described in The Cyberiad by Stanisław Lem.

That is to say, however much you pre-sort your data, there will still be a virtual infinity of them still left, all abiding your pre-set conditions, but all still uninteresting.

Can't a human draw a picture of Usain Bolt in the Greek Olympics? Wouldn't such a criticism also apply to human art?

I don’t understand you. The amount of human art, or indeed the possible maximum number possible of human artistic expressions (from the beginning of time until the end of the universe), is, I believe, far smaller than the pictures you would have left after your sorting filter.

I don't understand you either.

I don't know what you mean by a 'sorting filter', but what I'm talking about is an algorithm that takes >40Mb of data as input (pictures of Bolt and Athens for example) and uses this data to output ~40Mb of Bolt in Athens.

I mean human artists to be an analogy: humans can 'select' a picture of Athens Bolt without enumerating all others. I don't see why computers couldn't.

xkcd has a good take on this: http://what-if.xkcd.com/34/

It's discussing tweet variations, which is a much smaller set than the set of valid 5MB JPEG files, but even then it's rather a lot.

Also had a similar thought experiment, went and built it. Thanks for the other commentors as well I should link to this HN page in the README.

https://github.com/abemassry/crazip

Thanks for this. Have you considered using a GPU to speed up computations?

I thought about it, but according to the other commentors here a project like this wouldn't make sense. I was learning python at the time and I wanted to learn about multithreaded programming in python.

I could turn it into a learning project with no practical applications. It would be a good platform to learn about OpenCL, and see if the benchmarking times improve.

You're right, it is absolutely senseles. However, that doesn't mean you shouldn't take pride in your work.

Take a look at my Pythagorean cipher. It's idiotic and not practical, but it's just neat.

https://github.com/sp1d3rx/nonsense/blob/master/cipher.py

You may be interested in Borges's short story The Library of Babel.

http://hyperdiscordia.crywalt.com/library_of_babel.html

""The Library of Babel" (Spanish: La biblioteca de Babel) is a short story by Argentine author and librarian Jorge Luis Borges (1899–1986), conceiving of a universe in the form of a vast library containing all possible 410-page books of a certain format." (from wikipedia)

I immediately thought of the same thing. Personally, I like the idea of the Library of Babel more than, say, "all data is found in Pi". There's an important aspect to the story that really drives home the point, a hopelessness in the search as well as dealing with the nigh-incomprehensible scale. Well worth the read, as it brings the idea to a visceral level.

I'm at a loss for the story about indexing meaningfully the library, but I think I may have found a neat article on the Library: james.grimmelmann.net/files/Library.markdown

The identifier for any particular 'generated' photo is equivalent to the photo itself.

In reality there is no generation at all. It's just a blank slate with every bit unknown.

It's a perceptual problem, human don't understand randomness well.

I don't blame you, it's a common mistake, but you're fundamentally wrong in calling it generation. Generation of everything is 100% equivalent to generation of nothing. Creation requires decision, is equivalent to decision.

Here is an interesting article on a business plan is to copyright every conceivable image and make money on the royalties:

http://torrentfreak.com/copyright-apocalypse-trolls-attack-t...

>“Qentis aims to produce all possible combinations of text (and later on images and sound) and to copyright them,” Qentis’ Michael Marcovici told TorrentFreak.

>“Concerning text we try this in chunks of 400 word articles in English, German and Spanish. That would mean that we will hold the copyright to any text produced from now on and that it becomes impossible for anyone to circumvent Qentis when writing a text.”

>In terms of graphics, Qentis promotional material states that a subsidiary has already generated 3.23% of “all possible images” in the 1000×800 pixel format.

>“We are now generating images at a much faster pace and expect to complete 10 percent of all possible images by the end of 2015. At current projections, we will by 2020 generate every possible image in the 1000×800 pixel resolution,” the company claims.

Reply on news.ycombinator.com