First order founders are obsessed with GenAI — slapping image generation or a VLM with tools on problem XYZ. Second order founders are looking for defensible plays where they can create their own data flywheels. They are re-using the general techniques for building models for ignored areas like genetics or physics. Third order founders are seeing that the future of ML is not about massive data given to the machine by a human but data the machine gives to itself by theorizing and testing.
The smartest founders are grounding this theorizing and testing inside the real world to rapidly validate their ideas for creating auto-progression and the creation of self-curriculums.