“The debate over AI model architecture should be framed in terms of compute efficiency rather than whether a specific architecture like the Transformer can achieve AGI.”