GPT-5 was hyped as the best AI model ever, but users found it underperformed in real-world tests.
As heard by us
A clear look at how AI rankings are built, and how easy they are to game.
The episode frames AI reputation as a test design problem. The guest argues that models get judged against benchmarks built by labs or eval firms, then tuned to do well on those same measures once they are public.
Why you'd press play
Press play if you want the messy business of AI rankings explained without the hype.
Listen to the show on