Raw LLM Responses

Yeah openai and others have been leaning really really really hard into the "exponential" idea, that the step change between GPT-2 and GPT-3, and the step change between GPT-3 and GPT-4, suggests an exponential. Using really nebulous terms like "as smart as an average high schooler" (?) and "smart as a PhD researcher". Obviously that framing is ridiculously simplistic, but it's in our heads that that's how the creators of these systems are thinking of these step changes. So, expecting reasonably from what they said that GPT-5 would at least somewhat feel like a step change from GPT-4/4o. But in truth it's not even very empirically clear that that's the case. Benchmarks would suggest a marginal improvement, but even those can't really be trusted and don't seem to translate well to real world appilcations.