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.

I kind of gave up on this for a while but when Casey Newton and others started singing the praises of Glaze I thought, OK, let's try this. And Glaze pulled it off. I think part of the win here for them is that they have focused on one specific dev path (making MacOS apps) that suited me.

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!

But also, I was keenly aware that in building this tool, I was walking exactly the same road that any AI developer would if they wanted to snatch a mountain of Axios content for model training without, you know, asking.

This hit home particularly when I had to help Glaze devise workarounds for some of the safeguards Axios's team has put in place. The app has the user manually check the "human verification" box. Well, I'm a human! But yeah, that is my bot, not me, downloading the pages.

I mean, the same thing happens every time you read a web page, it's literally a "browser agent" that goes and gets the page for you so you can read it, you aren't walking over to the server and carrying the bits back to your screen yourself.

Yes, there's a difference in quantity between reading one page and 600. But I can't say I'm clear where exactly you draw a boundary between "human speed" and "bot speed" for page-loading.