Why GPT-5 was never going to impress you - Azeem Azhar's Exponential View Recap

Podcast: Azeem Azhar's Exponential View

Published: 2025-09-24

Duration: 7 min

Summary

GPT-5 was met with mixed emotions, largely due to two paradoxes: shifting goalposts and the negative space paradox. These paradoxes suggest that as AI advances, our expectations evolve, making it difficult for any new model to truly impress.

What Happened

GPT-5's release was marked by a sense of underwhelm, which is attributed to two historical paradoxes that affect our perception of technological progress. Azeem Azhar describes GPT-5 as evolutionary rather than revolutionary, highlighting that it was bound to fail in impressing us due to shifting expectations.

The first paradox discussed is the concept of shifting goalposts, a phenomenon observed since the 1970s. Azhar references Rodney Brooks, who noted that once a piece of AI's capabilities is understood, it stops being seen as magical and is relegated to mere computation.

The episode delves into the historical context of AI, recalling Alan Turing's foundational work in machine intelligence and the Turing Test as a benchmark. However, today's AI models like GPT-5 easily pass this test, leading us to constantly redefine what success in AI looks like.

The second paradox, the negative space paradox, is compared to the evolution of commercial flight. Initial awe at transatlantic flights quickly gave way to complaints about duration and comfort, mirroring how improvements in AI make its shortcomings more apparent.

Azhar explains that as AI models become more reliable and faster, like GPT-5, they still lack certain capabilities such as long-term memory and generalized intelligence. This leads to a situation where improvements highlight what is still missing.

He provides an example of AI's utility in automated workflows, noting that even a 1% error rate, though improved, can have significant impacts in complex processes. This subtlety makes it difficult for dramatic breakthroughs to be perceived as impressive.

Ultimately, Azhar argues that due to these paradoxes, we are unlikely to experience a single, transformative moment with AI as we did with ChatGPT. Instead, improvements will continue on a gradual curve, leaving us with a persistent sense of underwhelm as new models are released.

Key Insights

• The shifting goalposts paradox shows that once AI capabilities are understood, they are no longer seen as magical, reducing the perceived impact of advancements like GPT-5.
• AI models like GPT-5 have surpassed the Turing Test, leading to a continuous redefinition of what constitutes success in AI.
• The negative space paradox highlights that as AI models become more efficient, their remaining limitations, such as lack of long-term memory and generalized intelligence, become more apparent.
• Even with a reduced error rate, AI's impact on automated workflows can be significant, as a 1% error rate in complex processes can still have substantial consequences.