I can tell you from first hand experience that, since long before LLMs were invented, critical software supporting these industries is held together with duct tape and baling wire (and Excel). "Durable" does not mean "good". In my experience production code is often ugly, poorly abstracted, full of special cases and hacks, but most importantly it works.
----------------
Ok it can complete a line of python, it will never write a full, correct function.
Ok it can write full, correct functions but it will never write full working programs.
Ok it can write full working (disposable) programs but it will never write real, mission-critical (durable) code.
Why ?
In the 60s we thought we'd be living on the moon and mars "soon", flying cars were talked about seriously, that faster than sound jetliners were just around the corner, &c. Sometimes what's possible isn't desirable, or what's desirable isn't possible. No one went back on the moon, space programs are basically dead, planes are still as slow as they were in the 50s, a 747-8 from 2008 isn't much faster than a 707 from 1956
Never stop trying != succeed
This library (including the schema documentation) was largely written with the help of Claude, the AI model by Anthropic. Claude's output was thoroughly reviewed by Cloudflare engineers with careful attention paid to security and compliance with standards. Many improvements were made on the initial output, mostly again by prompting Claude (and reviewing the results). Check out the commit history to see how Claude was prompted and what code it produced.
"NOOOOOOOO!!!! You can't just use an LLM to write an auth library!"
"haha gpus go brrr"
In all seriousness, two months ago (January 2025), I (@kentonv) would have agreed. I was an AI skeptic. I thought LLMs were glorified Markov chain generators that didn't actually understand code and couldn't produce anything novel. I started this project on a lark, fully expecting the AI to produce terrible code for me to laugh at. And then, uh... the code actually looked pretty good. Not perfect, but I just told the AI to fix things, and it did. I was shocked.
To emphasize, this is not "vibe coded". Every line was thoroughly reviewed and cross-referenced with relevant RFCs, by security experts with previous experience with those RFCs. I was trying to validate my skepticism. I ended up proving myself wrong.
Again, please check out the commit history -- especially early commits -- to understand how this went.
It took an expert and a whole team of software engineer to code "with IA" something well defined, well known, and with already lots of implementation in other languages
Glorious indeed, we have created a new kind of printer ! (irony inside)
(in conclusion, the above post from cloudflare is an ad)
an expert using high powered autocomplete to successfully take a shortcut towards a well defined problem is not evidence of "AGI" or "the singularity"
belief in "the singularity" is a religious belief. you people sound like evangelicals talking about the goddamned rapture
To be fair, they did synthesize gold in 1984, "although the production cost is estimated to be a trillion times the market price of gold."