YouTuber PewDiePie reveals self-made ChatOS powered by Chinese Qwen models and Modded RTX 4090s

Felix Arvid Ulf Kjellberg, AKA PewDiePie, recently released a video on YouTube in which he revealed his big tech project built at home. The YouTuber has created his own AI service called ChatOS that uses Chinese open-source models with NVIDIA GPUs.

In the video, he mentioned how he leverages PCIe burification to essentially create a 10-GPU mini-datacenter, which was later utilized for a personalized AI service. His GPU rack was revealed to feature 8 NVIDIA RTX Ada GPUs and two blower-style RTX 4090s, which appear to be similar to those that are modded and widely used in China.

Felix decided to experiment with AI model hosting by hosting models like the Llama 70B, and the LLM ran successfully. Further, he created a web service to communicate with his local AI models, integrating functionalities such as web search, RAG, audio output, and memory. Baidu’s Qwen open-source model was also utilized to create a fully private and self-hosted system.

Multiple LLMs were run locally to see how they interacted with each other, nicknamed The Council and later The Swarm. The Council voted on a response to a prompt, to analyze whether it was the correct one, and the voting mechanism was operated by multiple AI instances.

By the end, the Council members began to strategically favor each other strategically, exhibiting a human-like behavior of colluding with the voting system.