The episode opens with a long discussion of DeepSeek, its V3 and R1 reasoning models, and why the release caused such a big reaction in AI circles and on Wall Street. Andrew says DeepSeek appears to have made real efficiency gains in training and hardware use, while Justin argues the market overreacted to the idea that less compute would be needed; both stress that the models do not mean chips or compute are suddenly unnecessary (L17-L17, L25-L25, L49-L49, L53-L57, L61-L65, L73-L77, L101-L105). The conversation then shifts to OpenAI's O3 and to a live, hands-on demo of generating simple games and 3D scenes with AI. Brian and Andrew iterate on a crude side-scrolling Mobius-strip game in CodePen, then experiment with A-Frame, a generated planetarium, and an explainer for radio telescopes, using the examples to argue that AI is becoming a practical tool for prototyping, brainstorming, and building educational or creative projects faster (L117-L117, L123-L145, L149-L181, L191-L209, L235-L241, L247-L253, L275-L289, L315-L317, L323-L333). Key topics DeepSeek efficiency gains and model optimization: Andrew describes DeepSeek as having made real, original optimization improvements, especially around data movement, compression, and training efficiency under chip export constraints. Uncertainty about data provenance and bootstrapping: The hosts note possible data contamination or use of model outputs, but they are careful to say those suspicions do not fully explain DeepSeek's success