> I'm speculating that much of the recent uptick in outages is simply the consequence of devs putting AI-written code into production without the same level of review/scrutiny that human code used to get.
The code they put up more than 3-6 months ago depending on the provider, I'd agree. But 6 months ago was roughly when the best LLMs got better at spotting issues than humans, and 3 months ago was roughly when open weight models passed the same threshold.
Not that this means everyone's using the best models, of course. Number of times I see academic research and it's "we used GPT-4o and it was bad" and I check the date and oh look one was submitted yesterday: https://arxiv.org/abs/2611.20753