Post

Here's a bit of a 🧵 with this one user's experience trying to get a little smidge of productivity out of generative AI coding. TLDR: it's taken three years, but (one of) the bots has finally accomplished a simple task for me, so chalk one up for AI.

I have largely been a gen-AI skeptic since ChatGPT's advent. We're always two years off from AGI or recursive self-improvement or superintelligence singularity, no matter when you start. It's like the whole industry closed its eyes and adopted the ludicrous Musk promise horizon.

But I also try to make sure my skepticism doesn't blind me to actual real-world advances — not benchmark BS but everyday accomplishments — and here is one.

I always wanted a local archive of the 600+ articles I wrote for Axios during my seven or so years working there. Axios' site has a nice author page but it's chunked into 5-stories-at-a-time and the prospect of manually copying those files always daunted me.

When people started claiming that AI was writing decent code in 2023, I began asking chatGPT and later Claude and other tools to do this job for me. The tools failed miserably for a long time.

They did keep getting better but there were always roadblocks. Recently it became clear that just writing code to crawl the Axios site and get the files was within reach. But doing so would still require me to set up a development environment, manage the code, run the debugging process, etc.

I could do that with effort. But the whole point here was to eliminate effort! If this was going to be a time-saver, AI would have to do more of the grunt-work itself.

At first the glaze agent made a series of mistakes that rendered the app useless. But its debugging was smart and easy. One by one, it took in my bug reports ("it's not getting past the cloudflare challenge page" or "we've got a full list of URLs but no body text") and fixed them.

It took 20 minutes to the first build and after about an hour of debugging I had my archive. I also had an app that I could "publish" and share with former colleagues if they wanted their own story archives. I haven't done that yet.

To me all this just confirms that the world of abundant bespoke personal software that Casey, Paul Ford and others have begun discussing is a real possibility. I find that exciting!