GPT 5.4 vs Gemini: Benchmarks, Codex, Excel

The podcast analyzes the performance of GPT-5.4 against other AI models like Gemini and Claude, particularly in the context of white-collar job automation. It highlights GPT-5.4's success in the GPT-Val benchmark, designed to measure AI's ability to perform real-world tasks, where it outperforms human experts 83% of the time. However, the discussion reveals that GPT-5.4 doesn't consistently lead across all benchmarks; Gemini 3.1 Pro Preview excels in areas like omniscience and reasoning. Karl Yeh joins the conversation, emphasizing the importance of testing AI models on specific use cases rather than relying solely on benchmarks. The hosts also explore the practical applications of these models, including their integration with tools like Excel and Google Sheets, and discuss potential privacy concerns related to devices like Meta Ray-Ban glasses.