On February 26, 2026, renowned YouTuber Felix “PewDiePie” Kjellberg made headlines with a provocative video. Titled “I Trained My Own AI… It beat ChatGPT,” Felix claimed that his newly developed AI model surpassed ChatGPT-4 and various other prominent AI systems. Right from the outset, he declared to his viewers, “The deed is done. I can finally return to this channel ’cause I have done what I said I was going to do. I trained my own AI model.”
He proceeded to share benchmark results that he asserted validated his claims. PewDiePie stated, “I ran the benchmarks, official AI benchmarks, and my model outperforms DeepSeek 2.5—a model much larger than mine. Facebook’s leading model, LLaMA 4 Maverick, was also defeated. Bang! Most crucially, my model surpasses ChatGPT-4, like, back in November or something.” The video sparked considerable discussion online, with fans questioning the seriousness of his comparisons.
Felix “PewDiePie” Kjellberg shares insights on Qwen 32B and the benchmark tests supporting his ChatGPT claims
As the video progressed, Felix clarified a critical point: he didn’t create a new AI from scratch. He explained, “I have not created my own AI. I have merely taken an AI model and trained it. It’s like stealing a child on the street instead of birthing one myself. It’s way more effective that way. Plus, it would cost millions in infrastructure, which I do not have yet.”
His project was based on Qwen 32B, a pre-existing model known for its coding prowess, but he aimed to enhance its ability in a specific coding format. “The model I used was Qwen 32B, which is already incredible at coding, but I needed it to excel in this format,” he noted. During testing, he focused on a benchmark called Ader Polyglot, which assesses coding capabilities across six languages.
He revealed that several leading AI models did not meet expectations on this test, stating that ChatGPT scored 18.2 percent initially, while Qwen 32B only managed 8 percent at first. The breakthrough came when he changed from the DIFF format to the HOL format, resulting in a score jump to 16 percent. He likened this shift to drawing a picture and adding a cloud without having to redraw the entire scene. This analogy emphasized his point that format is crucial and that minor technical tweaks can significantly alter outcomes.
PewDiePie’s venture reflects his growing interest in open-source AI. He had previously mentioned hosting large AI models, including OpenAI’s GPT-OSS 120B, on his own computer. The debate about whether his model genuinely “beat” ChatGPT remains open. However, his video undeniably ignited renewed discussions about how we evaluate and compare AI tools.
In conclusion, PewDiePie’s latest endeavor not only showcases his personal journey with AI technology but also highlights the ongoing conversations surrounding AI benchmarking and performance. As these discussions continue to evolve, the field of artificial intelligence is bound to witness even more intriguing developments.
