Given that you are an expert in the AI landscape, how different is coding and software engineering compared to a few years ago? What about 1 year ago?
I think that many of the things I currently do were possible with cursor 6months - 1 year ago, but would be unimaginable before that.
I also think the abilities of tool use by agents has improved a lot which improves the loops massively.
For instance, I had an issue with pulling a docker image recently due to disk space. I told cursor agent what command to run, and to keep trying to fix things until it responded with a 0 return code. It fixed it after a few iterations.
I think this type of thing was possible 6 months or even a year ago, but before that I don’t think the tooling was there yet. I could be wrong though.
I’m trying to get a sense of how fast things are changing.
My evaluation matches yours pretty closely. The thing that started working this year is agentic tool calling in the loop - there were systems that did this last year but they weren't quite good enough to get real work done. This year the models are better than the tooling - Claude Code, Cursor, Codex CLI etc - has improved to match.
Given that you are an expert in the AI landscape, how different is coding and software engineering compared to a few years ago? What about 1 year ago?
I think that many of the things I currently do were possible with cursor 6months - 1 year ago, but would be unimaginable before that.
I also think the abilities of tool use by agents has improved a lot which improves the loops massively.
For instance, I had an issue with pulling a docker image recently due to disk space. I told cursor agent what command to run, and to keep trying to fix things until it responded with a 0 return code. It fixed it after a few iterations.
I think this type of thing was possible 6 months or even a year ago, but before that I don’t think the tooling was there yet. I could be wrong though.
I’m trying to get a sense of how fast things are changing.
My evaluation matches yours pretty closely. The thing that started working this year is agentic tool calling in the loop - there were systems that did this last year but they weren't quite good enough to get real work done. This year the models are better than the tooling - Claude Code, Cursor, Codex CLI etc - has improved to match.