

Why GPT-5 was never going to impress you
36 snips Sep 24, 2025
Despite being the most advanced AI, GPT-5 left many feeling let down. The discussion highlights two paradoxes: shifting goalposts and negative space, which shape our expectations of new technology. Azeem explores how Turing's early benchmarks set unrealistic standards. As AI evolves, diminishing returns make previous capabilities seem less magical. The podcast delves into the concrete limits of current AI models and the compounding effects of minor errors. Ultimately, a true breakthrough in AI remains elusive, leaving listeners pondering the future.
AI Snips
Chapters
Transcript
Episode notes
Shifting Goalposts Dull Breakthroughs
- People have shifted the goalposts for AI so early tests like Turing no longer feel meaningful today.
- Progress that once seemed magical becomes dismissed as mere computation as capabilities become routine.
2014 Turing Test Controversy
- Eugene Goosman persuaded judges at Britain's Royal Institution in 2014 that his program was human, winning a Turing-style challenge.
- That early victory later got reframed as deception and a parlor trick as expectations rose.
Progress Makes Gaps More Visible
- Improvements in AI highlight remaining shortcomings, making gaps feel larger rather than smaller.
- Faster, more reliable models magnify contrasts like limited memory, lack of active learning, and imperfect generalization.