The article discusses the release of GPT-4o by OpenAI, comparing its speech synthesis to Google Duplex. However, the author suggests that if OpenAI had a more advanced model, GPT-5, they would have already presented it. The author also points out that GPT-4o is not significantly different from its predecessors, Turbo and GPT-4, and highlights some of the errors that have been reported, including reasoning errors and "hallucinations".
The author speculates that OpenAI's focus on new features may be due to their inability to produce the kind of capability advance that would align with predictions of "exponential improvement". The author concludes that the absence of a GPT-5 level model from OpenAI or any of their competitors suggests that we may have reached a phase of diminishing returns in this field.
Key takeaways:
The speech synthesis of GPT-4o is impressive, comparable to Google Duplex.
OpenAI does not have GPT-5 even after 14 months of trying.
There are still quirky errors being reported in GPT-4o, indicating room for improvement.
The lack of a GPT-5 level model from OpenAI or any competitors suggests we may have reached a phase of diminishing returns in AI development.
