On February 26, 2026, popular YouTuber Felix “PewDiePie” Kjellberg claimed in his video “I Trained My Own AI… It beat ChatGPT” that his trained AI model outperforms several advanced systems, including ChatGPT-4. PewDiePie emphasized that he trained an existing model, Qwen 32B, rather than creating one from scratch, likening it to enhancing a pre-existing child rather than birthing one. He revealed benchmark results from Ader Polyglot, showing that while ChatGPT scored 18.2%, his model initially achieved only 8%, which improved to 16% with a change in coding format. This experiment aligns with his growing interest in open-source AI. The video generated significant online discussion about AI comparison and performance metrics, raising questions about how such models are evaluated. PewDiePie’s insights highlight the importance of technical adjustments in AI training, contributing to the ongoing dialogue regarding the capabilities of modern AI technologies.
Source link
