AI Interview Challenge: Building a Classifier in Minecraft's Redstone
Description
A screenshot of a tweet by Scott Condron. The tweet describes an AI interview scenario. Below the tweet text is an in-game screenshot from Minecraft, showcasing an enormous and incredibly intricate machine built from colorful blocks. This machine is a Redstone computer, featuring layered circuits, complex wiring, and a large block-based display screen showing the number "3". The structure is vast, covering a large area in a sand-colored landscape under a clear blue sky. The meme humorously contrasts a standard AI interview task with a wildly unconventional and massively over-engineered solution. Redstone is Minecraft's equivalent of electrical circuitry, and dedicated players have built fully functional computers with it. The joke lies in the interviewer's initial disbelief followed by implied astonishment, as a candidate proposing to build an AI classifier in Redstone is either joking or a genius. For senior engineers, it's a nod to the concept of Turing completeness in unexpected systems and an appreciation for the monumental effort required to build complex logic from primitive building blocks
Comments
7Comment deleted
My first thought was 'that's insane,' but my second was 'I wonder if they used a block-based transformer architecture or a convolutional neural net for the logic gates.'
“Candidate’s explanation: ‘Gradient descent? Nah, I just route the error signal through a Redstone repeater loop until the torches stabilize - basically stochastic Creeper-dropout.’ At that point the panel quietly archived our entire MLOps roadmap
When your neural network has a 20-tick propagation delay per layer but at least the architecture is literally visible to stakeholders
When the candidate said they'd implement a neural network in Redstone, the interviewer realized they were about to witness either the most impressive demonstration of computational theory understanding or the longest 'compiling...' status in interview history. Turns out, when you ask for 'any framework,' some people take that literally enough to choose one where the garbage collector is a creeper explosion and the IDE is a pickaxe
When 'bring your own framework' means a 10k-block neural net running on observer clocks - no GPUs, just gravel traps for backprop
Finally, an interpretable model: every weight is a dust trail, backprop means relocating torches, and the throughput SLO is one prediction per loaded chunk
I said any language - they chose Redstone; now training runs at 20 Hz, inference latency is measured in chunks, and our only A100s are a hundred repeaters