UX is actually increasingly the bottleneck. Most of the top models are very good if you micromanage their context and prompts. But people aren't very good at that stuff.
Some of the desktop chat clients are turning into great productivity tools. I tried the Claude one last week and quickly went back to Chat GPT. Claude might be a better model for coding. But it's less effort to make Chat GPT do what I want at this point and it's kind of good enough for a lot of stuff. Every release it's getting better. It connects to my IDE automatically, it can look at the files I have open. It can patch those files (I actually disabled that because it's too slow for my taste), etc.
But most importantly, I can trigger all that with option+shift+1. I do this gazillions times per day. Mostly simple stuff with really short prompts, "check this" (file, selection, current line, etc.), fix that, what do you think about x, "address the FIXMEs/TODOs", "document this", etc.
I can ask other models the same questions and they'd get the job done. But then I have to do more work to give them the same context. Claude has a Github connect option, which is great. But unfortunately it's just a glorified file picker, which really sucks. I have files open in my editor, just look at those. I don't want to have to manually open files do that for me or specify what files to look at every time I go near the tool.
Chat GPT actually asked me yesterday whether it could add a different file than the one it was looking at. I said "yes" and it did. That's a great UX. Don't make me do work.
That's a good UX.
I use Gemini mainly because it's integrated into google's tools. So it's kind of there. And chat GPT for whatever reason does can not look at the browser window. But from a UX point of view, that kind of deep integration is what you want. You have this implicit shared context which is the thing you are looking at that you don't have to spell out anymore.
The UX of populating the context is the deciding factor in how useful models are at this point, not how well it solves pet benchmark questions or renders pelicans on bicycles.
I have good hopes for agentic coding tools progressing rapidly this year. The ones I've tried recently need a lot of work though. I keep going back to Chat GPT because it's just the quickest & easiest to use at this point.
Some of the desktop chat clients are turning into great productivity tools. I tried the Claude one last week and quickly went back to Chat GPT. Claude might be a better model for coding. But it's less effort to make Chat GPT do what I want at this point and it's kind of good enough for a lot of stuff. Every release it's getting better. It connects to my IDE automatically, it can look at the files I have open. It can patch those files (I actually disabled that because it's too slow for my taste), etc.
But most importantly, I can trigger all that with option+shift+1. I do this gazillions times per day. Mostly simple stuff with really short prompts, "check this" (file, selection, current line, etc.), fix that, what do you think about x, "address the FIXMEs/TODOs", "document this", etc.
I can ask other models the same questions and they'd get the job done. But then I have to do more work to give them the same context. Claude has a Github connect option, which is great. But unfortunately it's just a glorified file picker, which really sucks. I have files open in my editor, just look at those. I don't want to have to manually open files do that for me or specify what files to look at every time I go near the tool.
Chat GPT actually asked me yesterday whether it could add a different file than the one it was looking at. I said "yes" and it did. That's a great UX. Don't make me do work.
That's a good UX.
I use Gemini mainly because it's integrated into google's tools. So it's kind of there. And chat GPT for whatever reason does can not look at the browser window. But from a UX point of view, that kind of deep integration is what you want. You have this implicit shared context which is the thing you are looking at that you don't have to spell out anymore.
The UX of populating the context is the deciding factor in how useful models are at this point, not how well it solves pet benchmark questions or renders pelicans on bicycles.
I have good hopes for agentic coding tools progressing rapidly this year. The ones I've tried recently need a lot of work though. I keep going back to Chat GPT because it's just the quickest & easiest to use at this point.