The internet was graced over the weekend with “SmashorPassAI,” a website that rotates through a series of images of AI-generated women. Hitting “smash” or “pass” advances to the next image, and users can save images that they like.
“SmashOrPassAI is a recursively self-improving system. The more you swipe, the more it discovers your ‘type’” the site's creator, Emmet Halm, wrote on Twitter. “It then uses these insights to prompt itself and generate more AI images that it predicts you will like. The goal is to create the most addicting & personalized image recommendation system. V1 is as simple as possible.”
The site that went viral, not for its sharp recommender algorithm (which seemingly doesn’t exist) or its perfectly executed concept (which is very stupid and sexist) but because it’s the perfect example of what a lot of AI programming dudes think is a brilliant idea.
Halm, who according to his social media profiles just graduated from Harvard, tweeted that he’s simply in the arena trying stuff. “The goal is to create the most addicting & personalized image recommendation system. V1 is as simple as possible. Future versions trained on current data will enable even more personalized images & user interaction in image generation. Coming soon!”
I don’t know what’s going on at Harvard, but this little website Halm made is exactly like fellow alum Mark Zuckerberg’s hot-or-not application FaceMash from 20 years ago, but with AI-generated images of women in bikinis instead of his own classmates (thank God).
One of the (many) problems with SmashorPass is that the version he launched is not “recursively self improving” and doesn’t do much of what he said it does. It’s not actually a recommender algorithm, and beyond the pre-loaded images of AI generated women, the site is not using any kind of AI whatsoever. It’s a series of 187 images cycled at random while the user hits the “smash” or “pass” buttons.
@bridge_harris linked to the Google API in a reply to Halm to make this point—that it’s just a carousel of .png files—and he confirmed that to be the case.
“yeah v1 is just this — taking the data from tn to generate new images,” he replied. “Doing this manually rn, focused on shipping quickly.”
People caught onto this immediately. “I tried it for 15 minutes, looks like it's not recursively self-improving, but it's using same images repetitively,” @kumarkaushal_ said.
“Self-improving over time, not immediately yet. This is just a bare bones V1. Updates coming soon,” Halm replied.
The images themselves are also not that convincing AI-generated images. It’s easy to produce far more realistic images with easily accessible online tools.
Halm claimed on Twitter that the site hit 50,00 users overnight since posting it.
Several people pointed out that if one wanted to learn linear algebra that badly, they could use any other category to do so. “Want to get the most users possible to gather data,” Halm posted in response to someone making that observation. “10x easier to get ppl to swipe on hot images than plants. we’ll see if ppl like it. In the arena trying stuff.”
Wanting to learn how to code better and using images of women to do so has brought us deepfakes, many undressing apps, and for better or worse, Meta.
Rona Wang, a student at MIT in math and computer science, made “FriendorFoeAI” to lampoon SmashorPassAI. Like Halm’s site, it rotates through a series of AI generated images of men and asks users to decide whether the man is “friend” or “foe.”
“Upon seeing a woman, men decide immediately if they find her physically attractive; upon seeing a man, women must decide whether or not he’s likely to cause her physical harm,” Wang told 404 Media. “It works the same way as Smash or Pass in that it randomly shuffles pre-made images. No data is being collected.”
It would embarrass me, personally, to submit SmashorPassAI as a project for my community college Computing 201 class, let alone show it off on social media as someone who seems to want to be taken seriously as an app developer. That said, I have no doubt that this keeps a certain type of guy entertained for many minutes, like a dog with a puzzle treat. And Halm is already getting serious feedback: “The smashability of each photo does not seme[sic] to be increasing over time, which seems like a fundamental failure of the project,” @MotusRectus wrote. Interesting.
Halm did not reply to 404 Media’s requests for comment.