GPT-4

In ToTheMoon episodes GPT-4 is the reference point the hosts measure everything else against: the compute needed for the next generation, the level open models have reached, the plateau in language models and the confusion in OpenAI's lineup. It is rarely an episode's topic; it appears as the yardstick for Llama, Claude, GPT-4.5 and the promises made about GPT-5.

In 2024 GPT-4 marked the upper bound. In episode 001 the hosts said Russia could not build a GPT-4 for lack of hardware, and quoted Sam Altman's promise that GPT-5 would improve on GPT-4 as much as GPT-4 had on GPT-3. Episode 002 already called GPT-4 Turbo a fairly old model. In episode 003 the news that Llama 3 matched GPT-4 on several benchmarks cut both ways: the hosts no longer needed GPT-4 for their own tasks, but an open model at that level with its limits removed was unsettling. By episode 011 the talk was of a plateau: many GPT-4-class models, no real gain in a year.

GPT-4 then became a unit of measurement. In episode 019 the path to strong AI by 2030 was worked out in compute: GPT-4 was trained on twenty-five thousand GPUs, new clusters give six to eight times more, and a year-long run gives fifteen to twenty times. In episode 027, discussing the SB 1047 bill, the hosts stressed that simply adding parameters to GPT-4 would not lead to AGI, whatever that means. In 2025 the talk turned to the lineup: episode 048 described GPT-4.5 as a product with no obvious place and a price nearly a hundred times that of 4o, and in episode 061 the tangle of names showed how hard it is to pick a model for a task.