PewDiePie recently shared that he built his own AI program, and says it did better than ChatGPT on a test of coding skills. He explained the entire process in a new video, showing all the months of work, problems he ran into, and difficulties with the technology.
The project began as a way for the creator to learn, not to build an AI from the ground up. He didn’t create a new AI; instead, he improved an existing one by training it with specific data and coding tests.
PewDiePie explained he was working on a coding model designed to work well with AI coding tools. At first, it didn’t perform very well, falling behind other leading models. However, after several improvements to the training data and repeated practice, its performance steadily increased.
Benchmark claims and training process
In his video, PewDiePie tested how well different AI models – like DeepSeek, Meta’s Llama, and ChatGPT – performed using a coding challenge. He then shared the results.
He said his model initially scored around 8% on the standard test, then improved to 16% after some formatting changes. He further enhanced it with reasoning data and more training, and reported a peak score of 19.6% in one instance – briefly exceeding ChatGPT’s performance at that time.
He initially thought his results were good, but he later found that some of the data used to train the model was also included in the test questions. This mistake made the results unreliable, so he had to rebuild the model using a new dataset.
After being specifically trained on coding tasks, PewDiePie reported a significant improvement in the model’s performance. Initially, the score climbed to 36% once some testing problems were resolved, and then further increased to 39.1% after additional refinements were made.
The video explained some of the technical difficulties encountered during development, such as system crashes, overheating, and broken hardware. PewDiePie mentioned that one graphics card stopped working during the training process, and managing power usage was a constant challenge because the software required so much processing power.
He explained that the system was significantly customized and required constant repairs and rebuilding to continue the training process.
Even though the project faced some challenges, he explained that it was a valuable learning experience, especially in areas like machine learning processes, getting data ready, and building models.
When sharing the initial test results, PewDiePie pointed out that doing well on one test doesn’t necessarily mean the model is better overall. He’s planning to run more coding tests before deciding if he’ll make it available to the public.
He also pointed out that more recent models, like Qwen 3, now perform better on the same tests, so continued effort is necessary to remain competitive.
In the video’s conclusion, PewDiePie explained that the main goal of the project was to learn by trying things out and seeing what worked – and what didn’t. He mentioned he might keep working on the model, or start something new later on.
Read More
- All Golden Ball Locations in Yakuza Kiwami 3 & Dark Ties
- Gold Rate Forecast
- A Knight Of The Seven Kingdoms Season 1 Finale Song: ‘Sixteen Tons’ Explained
- Hollywood is using “bounty hunters” to track AI companies misusing IP
- These are the 25 best PlayStation 5 games
- Mario Tennis Fever Review: Game, Set, Match
- What time is the Single’s Inferno Season 5 reunion on Netflix?
- All Songs in Helluva Boss Season 2 Soundtrack Listed
- The MCU’s Mandarin Twist, Explained
- Best Zombie Movies (October 2025)
2026-02-27 23:49