If you read the entire internet multiple times over you would be the smartest person on Earth, so why is it that every frontier model has done that and you still have to read over a simple email ChatGPT writes?

Tune into this week’s episode to hear Maya’s answer: AI models are getting trained too inefficiently and Moonshot AI’s Kimi K2 is an example of how training can be more efficient. Kimi K2 is China’s latest open-source model, which was trained for agentic use cases. It’s doing better than anyone expected, and if you’re getting DeepSeek deja vu so are we.

The episode also dives into the AI browser wars and Priya gives her hot take on who she thinks has a right to win between Perplexity’s Comet, Dia by the Browser Company, and ChatGPT Agent.

Finally, the pod debriefs Lovable’s $200M Series A and the future of GPT wrappers.

00:00 AI browser wars

25:56 State of AI in China

52:18 Lovable raises a $200M Series A