“Dan Shipper theorizes that weaker AI models can often reproduce discoveries of frontier models if they are provided with the right conceptual hints.”