Teletext. Notepad. Twitter. A tractor. A pregnancy test. There have been few limits to the weird and wonderful ways enterprising Doom fans have found to play the seminal FPS over the past 30 years, but it could claim both the weirdest and the most wonderful crown. Someone teaches a bunch of lab-grown rat neurons to play Doom. Yes, their literal conscious existence is Doom. I told you it was weird.
YouTuber The Thought Emporium’s recent video details – with an emphasis on the details – their process of developing a rat neural network, with the ambition of eventually hooking the brain-like creation to a computer running Doom and teaching it to master the shooter. (Thanks, PC Gamer.)
While the YouTuber has experimented with human stem cell-derived neurons before, rat cortical neurons are apparently cheaper, easier to find (you can apparently just buy millions of neurons online, which I learned today), and are smart enough to learn Doom.
Neurons are placed in a set of small pot-shaped containers, where they are fed with a very specific solution to allow them to grow and crawl (!!!), gradually forming connected networks which can then be linked to a PC via an electrode network. (As a slightly disconcerting note, there is regular reference to a refrigerator-like machine called Meatcubator being used to help prepare neurons.)
The plan is to train the neural network to successfully detonate the cacodemons using electrical noise that the neurons “like” and “dislike”, rewarding them for killing enemies while punishing them for their deaths – effectively Pavlovian conditioning for the neurons. The video admits there’s no consensus on whether pleasant noise comes close to a treat or an orgasm, which is quite the thing to imagine while teaching a brain in a jar to play Doom.
While it might seem like overkill for a bunch of neurons in a dish to master a fast-paced first-person shooter, The Thought Emporium explains that Doom’s lack of top-down aiming means it can actually be simplified considerably – enough that it becomes a very basic set of instructions for turning left and right, forward and shooting as delivered by “yes” and “no” commands to the neurons.
“The reason we chose Doom is that it’s a complex enough task to be impressive, but it can be deconstructed into a simple signal problem if you set it up correctly,” they explain. “Doom only looks 3D. In reality[,] the character is just an arrow, the enemies are circles, the hazards are just shapes, all on a 2D surface. The game just renders it to look 3D.
The first video in the planned series depicts the growth of neurons, but they have yet to connect and start learning Doom. The next steps require finding a way to receive signals outside of the neural network, as neurons can begin to control Doom and learn to improve. The team also hopes to find a way to grow neurons from human skin samples to build a stable and affordable supply of neurons. (They add that mixing in a rat would also work, but they’re not about to.)
Whatever the end result of this experiment, it’s hard to deny that it’s a fascinating and slightly terrifying raising of the bar for ways to play Doom.
