You Say Potato, I Say GLaDOS

How many times have you thought to yourself: “I really wish I had an artificially superintelligent computer system that was low-key trying to kill me”? It is probably too many times to count if you are like most people. Unfortunately, building evil, superintelligent machines is out of reach for nearly everyone. YouTuber Dave's Armoury found this out firsthand when building a malevolent GLaDOS robot inspired by the video game Portal. Between the robot arm, pricey edge computing platform, and everything else that went into it, it cost around $20,000.

The goal is for the robot to try to kill you, not your credit card bill, so Dave's Armoury recently reworked his version of GLaDOS to make it more accessible. The idea came from Portal 2, in which the core of GLaDOS takes up residence in a potato, using it as a power source. That gets rid of the robotic arm right away, which saves a lot of cash. But there was still some more slicing and dicing to do to make this project accessible to the average hobbyist that just wants to experiment with AI.

To work anywhere, the device must be able to work without an internet connection. That means cloud-based services cannot be used. Accordingly, the initial version of GLaDOS was powered by a $2,000 edge computing system. But Dave's Armoury thought it might get along well enough with a $250 NVIDIA Jetson Orin Nano. However, it does have less compute power and memory available, so some other cuts had to be made.

For starters, NVIDIA’s Riva had been used for text-to-speech services. It requires too much memory for the new hardware, however, so it was replaced by Piper, which can even run comfortably on a Raspberry Pi. GLaDOS is not GLaDOS without the right voice, so a large number of audio files were scraped from the Portal wiki and used to train the Piper model to have the right sound.

GLaDOS also needs to hear what the user is saying so that it can respond appropriately, so for this purpose Vosk was utilized. It did have a higher error rate than Riva in testing, but it still got the job done well enough for a beginner project.

The final major component was the large language model (LLM). The previous model was too large, so Dave's Armoury selected a Llama LLM with three billion parameters. A retrieval-augmented generation approach was used to supplement the model with additional information so that it would be aware of the Portal universe, then an elaborate system prompt was constructed to give GLaDOS her trademark personality.

To finish things off, the hardware, along with a power bank, was fitted into a 3D-printed potato that was painted to add realism. Some RGB LEDs were also added into the build to indicate the state of the system (listening, thinking, etc.), and just because they are awesome.

Take a look at the video to see the full build process for this scaled-down version of GLaDOS. At long last, killer robots are finally within our grasp!