I think the experience of the gpt4/gemini2/claude2-3 era of models was widely variable depending on the use case and how much reference material there was on the internet. I'm told it was very good at producing react code for example, but my experience was it often failed to handle Rust or Kotlin type checking and failed pretty often (in less immediately detectable ways) on PHP. So looking at the trajectory from the initial copilot (could auto compete fast inverse square root and method level problems that you could also just google), to GPT4 (which from my experience, still couldn't code) and hearing that the training data by that point was "most of the internet" it wasn't that clear to me that it would reach the point it has today.

Obviously the models themselves have trended bigger since which has helped and also just the tooling and harnesses around it and the models being trained for that use case has achieved a lot since that q4 2025 window which is about the first time it became functional for me.