What number of occasions have you ever thought to your self: “I actually want I had an artificially superintelligent pc system that was low-key attempting to kill me”? It’s in all probability too many occasions to rely if you’re like most individuals. Sadly, constructing evil, superintelligent machines is out of attain for almost everybody. YouTuber Dave’s Armoury discovered this out firsthand when constructing a malevolent GLaDOS robotic impressed by the online game Portal. Between the robotic arm, expensive edge computing platform, and every part else that went into it, it price round $20,000.
The aim is for the robotic to attempt to kill you, not your bank card invoice, so Dave’s Armoury not too long ago reworked his model of GLaDOS to make it extra accessible. The thought got here from Portal 2, wherein the core of GLaDOS takes up residence in a potato, utilizing it as an influence supply. That eliminates the robotic arm immediately, which saves a whole lot of money. However there was nonetheless some extra slicing and dicing to do to make this challenge accessible to the typical hobbyist that simply needs to experiment with AI.
A glance contained in the case (📷: Dave’s Armoury)
To work anyplace, the machine should be capable of work with out an web connection. Which means cloud-based companies can’t be used. Accordingly, the preliminary model of GLaDOS was powered by a $2,000 edge computing system. However Dave’s Armoury thought it would get alongside nicely sufficient with a $250 NVIDIA Jetson Orin Nano. Nonetheless, it does have much less compute energy and reminiscence accessible, so another cuts needed to be made.
For starters, NVIDIA’s Riva had been used for text-to-speech companies. It requires an excessive amount of reminiscence for the brand new {hardware}, nonetheless, so it was changed by Piper, which may even run comfortably on a Raspberry Pi. GLaDOS shouldn’t be GLaDOS with out the precise voice, so numerous audio information had been scraped from the Portal wiki and used to coach the Piper mannequin to have the precise sound.
GLaDOS additionally wants to listen to what the person is saying in order that it will possibly reply appropriately, so for this function Vosk was utilized. It did have the next error price than Riva in testing, but it surely nonetheless acquired the job finished nicely sufficient for a newbie challenge.
Wiring up the Jetson (📷: Dave’s Armoury)
The ultimate main element was the massive language mannequin (LLM). The earlier mannequin was too giant, so Dave’s Armoury chosen a Llama LLM with three billion parameters. A retrieval-augmented era strategy was used to complement the mannequin with extra info in order that it could pay attention to the Portal universe, then an elaborate system immediate was constructed to present GLaDOS her trademark persona.
To complete issues off, the {hardware}, together with an influence financial institution, was fitted right into a 3D-printed potato that was painted so as to add realism. Some RGB LEDs had been additionally added into the construct to point the state of the system (listening, considering, and so on.), and simply because they’re superior.
Check out the video to see the complete construct course of for this scaled-down model of GLaDOS. In the end, killer robots are lastly inside our grasp!
