I’ve decided to finally hop on the bandwagon and give the sensationally popular Squid Game a watch. Having finished the series last night, I can appreciate where all the hype was coming from. As a fan of similar death game works in the past, such as the archetypical Battle Royale and the terribly schlocky anime King’s Game, this live action Korean take on the genre was an enjoyable experience from start to finish.
In the show, a heavily indebted gambling addict named Seong Gi-hun winds up invited to a mysterious event after winning 100,000 Won from a complete stranger. He’s faced with the risk of losing his daughter, with his ex-wife planning to move them abroad, and him unable to reclaim custody if he can’t prove himself able to support his daughter financially. With this relationship on the line, he decides to forsake the twin notions of stranger danger and things sounding too good to be true, and enters the contest. Only after the first game starts does he start to realise his mistake, when it immediately becomes obvious that being ‘eliminated’ for losing is not a euphemism.
This brings us to Young-hee, the four meter tall robot who serves as referee for the first game. Playing ‘Red Light, Green Light’, the objective is to run to the other side of the arena when Young-hee is facing away. If a player slips up and moves while she is looking, sniper fire takes them out.
There’s two components to this AI I’m going to discuss. The first is Young-hee herself, who acts as a complex motion sensor taking advantage of computer vision. The second is the sniper rifles enacting her judgement. No humans are seen firing them, so I’m going to assume they’re autonomous in targeting and firing for the sake of discussion.
We see a first-person perspective of how Young-hee spots unfortunate moving victims. Her eyes act as cameras, scanning the players for signs of motion. We can assume each works independently, as they are shown to be de-synced from each other.
Detecting motion itself is perfectly plausible. All Young-hee needs to do is identify what is a human in each image she sees, known as a frame, then compare it with a later frame to see if that human has moved. She can do this by storing the locations of each player in each frame, compare the pixels that make up each human image in subsequent frames, and look for big enough changes. If so, bye bye player.
Young-hee has to monitor 456 players, and we’ll make a conservative estimate that she records 60 frames per second. She also has to distinguish between the contestants to detect individual movement at a very long distance, including where players partially overlap. From this, we can make two assumptions.
First, we might assume the resolution of each camera must be excellent. Otherwise players would blur together, especially at a distance. Fortunately (or unfortunately if you’re playing the game), we live in the world of 8K cameras, with 8K UHD recording 7680 x 4320 pixels in every frame. At 33 million pixels, we’ll assume that’s enough to track each player sufficiently well, even at a distance. The fact that the cameras are shown zooming makes this less of an issue, though it raises a question of whether another AI component is judging whether she needs to zoom in or not to make a decision.
In practice, the view we get of Young-hee’s camera suggests a lower resolution than even 4K, unless it’s just my TV being too low res. Yet she’s still able to pick out the individual players well enough to tag who’s moving and who’s standing still. So what’s going on?
It turns out that AI’s ability to pick out targets is really quite powerful, even on lower resolution cameras. Here’s an example of street camera technology picking out individuals, even when they’re overlapping with each other. We call this technique of identifying specific regions from images using AI Image segmentation. In this case, we’re segmenting out the humans from the background.
Our second assumption is that Young-hee must have an enormous processor in her big ol’ doll head. Not only does she have to compare every single frame to see if someone’s moved, but she also needs to separate out every player in every frame while storing their current and previous positions. The first calculation, comparing frames, is relatively simple and proportionate to the camera’s resolution, but needs to happen every 1/60th of a second. The latter, depending on the algorithm, can cost much more and has to happen at the same pace.
This could imply a clever trick by the developers of the game to ease calculations. Each player wears a teal tracksuit, which stands out quite distinctly from the naturally coloured arena. All Young-hee has to do is compare the colour of each pixel against a stored shade for the tracksuit, and she can have a pretty good idea which pixels represent a tracksuit cladded human. This is further backed up by the fact that Young-hee never shoots players for talking, suggesting she isn’t watching their faces very closely.
This could be a potential way to cheat the game. If Young-hee is looking for a certain shade, then stripping out of your tracksuit could make you invisible to her. Of course, you have to managing stripping while her head is turned, hope your hair or any underwear doesn’t give you away, and forsake your modesty in front of 455 people in the process, but it’s an idea!
Sadly image segmentation is not an area I’m very well versed in, so this is about the limits of my speculating. But at the very least, I can agree that all the technology Young-hee is running on is well within the realm of AI that already exists.
While it’s possible that the snipers are all manned with some very skilled shots, it’s more likely and interesting in this case to consider them robot snipers. Something like this falls under the category of a Lethal Autonomous Weapon (or LAW). This controversial technology takes the human out of the picture when it comes to warfare, designed to seek and destroy its target without a human pilot. Arguments have been made for and against LAWs.
Proponents claim that they reduce the risk of things going awry on the battlefield due to factors such as stress and fatigue. They would minimise the human presence and hence danger to individual soldiers, and that by streamlining the war process it can be resolved more effectively with lowers costs. Ethicists have argued that may even be preferable to human soldiers, particularly due to lacking a need for self-preservation that could impair judgement.
Critics cite how the removal of humans lowers the cost of war, making it more enticing to aggressors. They also argue that these weapons can cause more destruction than humans could by themselves, and the they pose a great risk if they act in unpredictable or uncontrolled ways. Ethicists also highlight the question of accountability with regards to actions taken by LAWs: who is responsible for who the robot kills?
As technology goes, automatic sniper rifles are perfectly plausible. Acting as a sniper requires considering a wide range of variables, such as wind speed and barometric pressure. The role is frequently assisted by a second person, known as a spotter, who records the outcome of the shot and reports this back to the sniper in case they need to readjust.
A computer, in comparison, can perform both of these roles by itself. With the required sensors easy to construct in the dedicated arena, and the targeting mathematics performed accurately in fractions of seconds, it is simply a case of teaching the gun how to properly position then shoot. With dozens if not hundreds of players being mowed down in seconds, autonomous snipers make much more sense than humans, and also highlight the destructive potential of LAWs.
South Korea even gives us an existing example of an LAW in action, already employing automatic sentry turrets to guard its borders. The use of this technology in this context is not without controversy, of course, but we’ll avoid taking any definitive sides in this post.
Despite the terrifying might of Young-hee, nothing about the technology backing her or her barrage of sniper fire is implausible. The high level of her performance definitely requires some cutting edge technology however, so we’re unlikely to see anyone hosting a proper Squid Game in their garden any time soon. Yes, people have posted plenty of their own recreations, but fortunately these tend to be of the non-lethal, AI-free variety.
What do you think? Am I overstimating how powerful modern technology can be, relative to our killer doll’s modus operandi? Have any useful information on image segmentation? Or maybe you’d like to leave a few remarks on lethal robots? Share your opinions and views in the comments below! And for the benefit of people who haven’t seen the show yet please remember: no spoilers!