The LLM scrapers started to reliably knock nóssa offline (even with Anubis difficulty increased), so I had to implement Measures :( Logged-out users can view the root page of any repository, but can’t navigate to any blob, tree or commit. (You can still clone them fine, though!)

That seemed like a pity since I like to deep-link a lot in posts and emails and that kind of thing, so I added a “signed URL” feature. This link to the code responsible is an example!

It still seemed a shame: registration isn’t open on the instance, so in practice no-one can navigate repositories but me and Annie. For a long time I wanted to support “remote” users from e.g. Forgejo instances for eventual coordination activities like issue tracking, and so now seemed like the time: you can login with Codeberg, GitLab, or GitHub, and an account is created automatically if it’s not linked to an existing one, with no extra steps. (Shout out to Überauth and OIDC.) To prevent abuse new accounts are given “viewer” status and can’t create new repositories or have public profiles.

This took a couple hours of a Sunday evening, mostly debugging semi-arcane “invalid redirect URI” responses¹, and I’m really happy it’s done.

1. I needed to add Plug.SSL to the stack to translate XF* headers in the %Plug.Conn{}; Überauth was reconstructing “redirected to” URIs with the wrong scheme/port and complaining the redirect was made incorrectly. ↩

Cross-post from lottia notes.

The garbage just keeps getting worse, and something I’m struggling with now is that, increasingly, people are fooled by it! I stumble upon what was clearly one or two sentences fed into Claude and people are in the comments calling it the Best Article They’ve Ever Read On The Subject; worse, the comments are often by very (extremely) technically capable people, often those very disposed against LLMs! I don’t get it. They just managed to read several thousand words of nothing and, presumably(?) due to a lack of engagement with LLMs, they have no immune system for it. God help them if they actually try one once; perhaps these are the same people who get “one-shotted”.

A few months ago the reaction to these pieces was almost completely negative and the posts were buried quickly, but since the same updates landing that made these models quite reliable for coding tasks — which the anti- crowd by and large still do not think possible — the bullshit has really started to drown out the human-written stuff, with plenty of the perpetrators even full-on using them in the comments section of places known to be hostile to LLMs, and yet people’s bullshit detectors have fallen silent. It’s all a big yikes.

Senior Backend Engineer (Ruby on Rails), Plan: Knowledge