Summary
Theo of t3.gg reacts to a video from PewDiePie, who has moved from gaming content into building his own AI tools. The project, called Ajax, is a fine-tuned local model designed to browse the web, run private searches, write emails, and use a calendar, tuned for a specific agent harness and released in an uncensored form.
The video follows PewDiePie’s attempts to train a model of his own, including his effort to crowdsource training data through a dedicated submission website and his surprise that almost nobody was willing to contribute, even with anonymity and tutorials. It also covers his attempt to distill knowledge from a larger frontier model and the obstacle that OpenAI hides its reasoning tokens, which limits what can be learned from its outputs.
Theo adds commentary throughout on why open-weight models often lag frontier models, particularly in sub-agent orchestration, where training data is scarce because it cannot simply be scraped from the internet. Viewers interested in fine-tuning, distillation, local AI, and the practical challenges of data collection for model training will find a lively, accessible look at what an ambitious hobbyist project involves.
📺 Source: Theo – t3․gg · Published November 05, 2026
🏷️ Format: Reaction
