>Firstly, lots of people ask if it's an algorithm or something. It's not! Just me searchin google.
>Secondly, a lot of people assume this blog is therefore made predominantly by using the "visually similar" function on Google Image Search, which is a totally reasonable thing to assume. While I definitely employ that function in my attempts to search thoroughly (and love it for it's own beautiful results), surprisingly little of the piece is actually constructed using it! Visually Similar appears to employ a logarithm based mainly on color percentages in an image, and as I'm Google is based more often in conceptual similarities than color-wash similarities, my searching is almost entirely relient on keywords rather than searching by image. Of course, there are times when Visually Similar has helped with a transition or section here or there, but overall, it's not the way I work.
> Firstly, lots of people ask if it's an algorithm or something. It's not! Just me searchin google.
Love that...I sometimes worry we're getting to a point where those who haven't programmed before don't realize they can do things without apps. Such as making a data visualization by drawing. Or math.
Yes, the transition points were so incredible, and downright hilarious. Especially one with an explosion within a TV screen transitioning into actual sets of explosions. Makes a lot of sense now.
Aww man, I thought this really was pure algorithms, not manual curation. It is still a wonderful project but my awe is gone. Would be great to see it as a real bot.
When I read the title "I'm Google - ..." a very small part of me was really, really excited that maybe, just maybe, some system within Google became self aware, learnt english, and wrote a blog declaring its identity.
Man, you've made my day. I've always imagined the first PR about transient AI as "Dr. has applied for a patent for Y" or "How can Siri answer this, although she only knows that". But the idea of a system becoming transient and Google employees not knowing where it is stored inside their datacenters would be plausible. Assuming AI is possible.
Hopefully the first thing this AI will declare is whether it stores data on its EU server.
I am oddly reminded of Pollard's rho algorithim[1], you start with a base image and only choose the first image that comes up in the result, in the end we will eventually return to an image that we have already traversed. I wish someone could find the base images that result in the shortest and longest cycles.
My guess is the shortest would usually be the very next one since Google Image is good at finding the same image file in different websites. There would have to be a constrain for how close to the last image the next one is or you'd end up in a one image loop.
Some of the break rooms where I work have a computer you can do searches on. There are some Google image searches that just shine like diamonds. Most of them are somehow PBS related for some reason. "Bob Ross" "Fred Rogers" "Ernie and Bert", etc. I started an odd little game of people leaving google image searches on the break room machines.
Interesting. It seems to alternate between being "stuck" in certain conceptual valleys, and jumping between different concepts quickly. Perhaps an effect more similar to the op could be achieved by enforcing some sort of smoothness and similarity constraints.
I'm assuming porn is filtered out. Given that porn represents the vast majority of Internet content, it would be curious to see how quickly a sexually explicit image would pop up, and how funny the connection with the previous non sexual image would be.
It seems that according to modern estimates this thing about porn being the greatest part of the Internet is no way close to reality and was more of just a legend that got told over and over and ended up being accepted for truth.
But honestly I myself don't know and can't be bothered to investigate more. I just know there's enough of it :)
"Back in the days" image morphing software was somewhat popular. It could be fun project to run these images through some automated tool to create a continuous video of one image morphing to another then to third and so on.
You might not need to morph it, just play it back at 24 fps and the results might be quite interesting. Essentially the hard work of interpolating frames has been taken care of by choosing a slow progression of one image to the next.
I love how on Google when I do an image search on myself I can go from Brad Pitt one minute, to an old man the next.
That's not a complaint though. The technology is always improving and I live in constant wonder of the things people can create from nothing :)
A few scrolls down there is "This video was removed by Youtube for privacy reasons".. does the related image algorithm still use what image was there, for calculating next image?
I can't find where the about link is, but here's the about text i found on the source code:
I’m Google is an ongoing tumblr blog in which batches of images and videos that I cull from the internet are compiled into a long stream-of-consciousness. Both the searching and arranging processes are done manually. The batches move seamlessly from one subject to the next based on similarities in form, composition, color, and theme. This results visually in a colorful grid that slowly changes as the viewer scrolls through it. Images of houses being demolished transition into images of buildings on fire, to forest fires, to billowing smoke, to geysers, to bursting fire hydrants, to fire hoses, to spools of thread. The site is constantly updated week after week, batch by batch, sometimes in bursts, sometimes very slowly. <br><br>
The blog came out of my natural tendency to spend long hours obsessing over Google Image searches, collecting photos I found beautiful and storing them by theme. Often the images that interest me are of industrial or municipal materials or everyday photo snapshots. I do not select images or videos that appear to be intentionally artistic. Happily, the process of researching various themes in this way has lead to unintentionally learning about topics I might never have otherwise, including structural drying, bale feeders, B2P, VAWTs, screw turbines, the cleveland pack, and powder coating.
I feel that my experience wandering through Google Image Search and YouTube hunting for obscure information and encountering unexpected results is a very common one. My blog serves as a visual representation of this phenomenon. This ability to endlessly drift from one topic to the next is the inherently fascinating quality that makes the internet so amazing.
A NOTE ON THE PROCESS
Just wanted to add a note on how I make this blog, as I have seen people wonder the same couple things frequently.
Firstly, lots of people ask if it's a algorithm or something. It's not! Just me searchin google.
Secondly, a lot of people assume this blog is therefore made predominantly by using the "visually similar" function on Google Image Search, which is a totally reasonable thing to assume. While I definitely use that function in my attempts to search thoroughly (and love it for it's own beautiful results), surprisingly little of the piece is actually constructed using it. Visually Similar appears to employ an algorithm based mainly on color percentages in an image, and as I'm Google is based more often in conceptual similarities than color-wash similarities, my searching is almost entirely relient on keywords rather than searching by image. Of course, there are times when Visually Similar has helped with a transition or section here or there, but overall, it's not the way I work.
It is a wonderful, thoughtful project, but it's an art project.
While that's great, I don't come to Hacker News for the latest interesting art project.
I think like many people that clicked the link and scrolled through the images, we made the assumption that these were put together using some algorithm that mined Google's 'visually similar' search and then somehow found the best results in order to create a set of images that 'flowed'.
It's a little disappointing to learn otherwise, but it's now kicked off a train of thought about how I might automate something similar. :-)
The relevant bit:
>Firstly, lots of people ask if it's an algorithm or something. It's not! Just me searchin google.
>Secondly, a lot of people assume this blog is therefore made predominantly by using the "visually similar" function on Google Image Search, which is a totally reasonable thing to assume. While I definitely employ that function in my attempts to search thoroughly (and love it for it's own beautiful results), surprisingly little of the piece is actually constructed using it! Visually Similar appears to employ a logarithm based mainly on color percentages in an image, and as I'm Google is based more often in conceptual similarities than color-wash similarities, my searching is almost entirely relient on keywords rather than searching by image. Of course, there are times when Visually Similar has helped with a transition or section here or there, but overall, it's not the way I work.