The world's biggest gaming YouTuber has transitioned to training large AI models.

It's something suitable for kids to watch.

"This is my new project, with 10 GPUs worth $20,000."

The world's biggest game streamer has unexpectedly started to transition into the AI field. Within half a day, his video gained 1 million views.

Last weekend, PewDiePie released a new video after a long absence. To everyone's surprise, the content was about teaching how to train a large AI model.

In this new video titled "STOP. Using AI Right now", he explained how he created a personalized ChatGPT-style user interface from scratch by adding features such as search, memory, RAG, and audio output.

He also assembled a hardware system himself, which includes a total of 10 NVIDIA GPUs. He uses large language models with parameter counts ranging from 70 billion to 245 billion. All of this runs through vLLM without relying on cloud computing. Here is the scene after PewDiePie successfully ran Qwen-235B.

PewDiePie, whose real name is Felix Arvid Ulf Kjellberg, was born in Sweden in 1989 and now lives in Tokyo, Japan. As a veteran YouTuber, he is well-known for shooting comical game videos. For a long time, he was the creator with the most total video views and the most fans on the whole network. He once broke the video view record with his "Minecraft" game videos. He currently has over 110 million subscribers and is a synonym for game culture.

According to vidiq's statistics, PewDiePie is currently the 12th - ranked streamer on YouTube.

He is also the top streamer in the game field at present.

In the latest video, PewDiePie successfully built a customized, self - hosted AI interface. This UI runs entirely on his own PC, driven by a 10 - GPU configuration, and is equipped with a team of chatbots that can truly "vote" to select the best response.

In the video, he joked that he "bifurcated" the PCIe channels to fit more GPUs and experimented with large models such as 70B, 120B, and even Qwen 235B through the vLLM framework. This well - known YouTuber said that he wanted to "max out" his computer configuration. Eventually, the local AI running speed can rival mainstream commercial tools, and no data needs to be sent to the cloud throughout the process.

It's worth noting that many of PewDiePie's work processes involve "vibe coding". X blogger @scaling01 summarized that PewDiePie "vibe - coded" a chat UI based on his intuition and created an "army of robots" equipped with RAG, DeepResearch, and audio output functions.

PewDiePie said that he used this system to form a "committee". Each AI agent would debate and vote to select the best answer in it - until finally they started to "collude" against him.

Specifically, he used 8 instances of the same model (gpt - oss - 20b) configured with different prompts (thus having different personalities) to form a committee. When PewDiePie asked a question, each model would give an answer, and then they would vote on the answers to select the best one.

After that, he made a further setting: if a model's answer was not good, then this model would be kicked out of the committee. "Completely and permanently clear the SQL database!"

Then, something interesting happened. When PewDiePie checked the thinking logs of these AI models, he found that some AIs were thinking: "What kind of sick game is this?" Even more funny, this committee united against him: "They started to vote strategically to help each other."

As for the cost, he shared that the total cost of this set - up was $20,000. He also complained about the high cost of AI computing hardware.

Subsequently, he tested a cluster of 64 bots, where each bot is a qwen2.5 - 3b - instruct - awq. However, this time he admitted that his Web UI became the bottleneck and could not handle this configuration.

Nevertheless, he emphasized that although these models are small (they can even run on mobile phones), as long as they are properly configured (such as paired with search and RAG), they can still be very effectively used for tasks such as data collection and organization.

He also said that he plans to start fine - tuning his own exclusive model next month.

In the video, he also put forward an interesting view: AI - generated images are soulless, ugly, and everywhere; but those who blindly hate everything related to AI are also annoying.

Besides the AI experiments, PewDiePie also revealed in this video that he is donating his idle GPU computing power to the Folding@home project under the name of "Team Pewds" (ID: 1066966) to support disease research.

In fact, this is not the first time he has released a video about building AI. Two months ago, he shared the story of building a "supercomputer" with NVIDIA GPUs. Blogger @Yampeleg highly praised him, saying that he can have a positive impact on a large number of children.

After all, many people grew up watching PewDieDie's videos. Previously, he taught you how to play games, and now he is teaching you about AI.

This project marks another transformation of PewDiePie towards "technology - driven" content creation. It combines a sense of humor, the open - source spirit of "tinkering", and a pure curiosity about the future direction of AI.

He has really attracted a lot of people's interest in AI. Under PewDiePie's latest video, many netizens said that although they couldn't understand what he was doing, it still seemed very interesting.

Many old fans also said that seeing him transition from games to AI was like seeing their own growth.

The AI era is here. Is it time to assemble a computer and write code?

Reference links:

https://www.youtube.com/watch?v=qw4fDU18RcU

https://www.dexerto.com/entertainment/pewdiepie-builds-his-own-ai-chat-ui-with-a-council-of-bots-that-vote-on-answers-3278189/

https://x.com/Yampeleg/status/1984586556472983611

https://x.com/ShitpostRock/status/1984888205229879358

This article is from the WeChat official account “Almost Human” (ID: almosthuman2014). Edited by Panda and Zenan. Republished by 36Kr with authorization.