[17:03:41] #tech-ops os151 is not a happy chap. [17:04:01] Same with os201 and os191 as well. [17:07:39] It has been going on for more than a week I think. Log spam overloading greylog. It was initially wikibase but once that is patched deprecations warnings are generating way too much logs. [17:09:05] Ah. [21:14:38] [1/3] I was looking at [WO's T&S report](https://meta.wikioasis.org/wiki/Trust_and_Safety/Transparency/Q3_2026) by @zipppee and I found the use of classifiers interesting. It's definitely a good tool for this particular use case, though the rate of false positives seems pretty high. [21:14:38] [2/3] This might also be useful for CVT, with the AbuseFilter -> Classifier -> CVT feed route being less prone to false positives than before. [21:14:39] [3/3] Should be possible to deploy something like [laya](https://huggingface.co/convaiinnovations/laya) to the spare LLM server. It has 0.4b params which is nothing compared with the gpt-oss-20b we tried to deploy before. [21:17:39] the classifier is good for finding things that are bad [21:17:45] it doesn't know whether it's bad enough for T&S [21:18:22] I imagine it'd be useful for CVU [21:18:57] It'd be nice to see how many cases were actually upheld [21:19:41] And https://wikioasis.org/fandom also looks like an interesting idea. It seems to rely on a database dump instead of the WG grabber? It does make migrating much more enticing which is good. [21:19:50] Cause it's really hard to map how 2086 reports end up at less than 50 people acted on [21:20:00] Is it that 1950 reports were not useful [21:20:12] Or that one person uploaded all 300 CSAM images [21:20:16] damn the AI slop goes hard [21:21:40] This is basically what Cook has been suggesting. Not the make web pages with AI part, but making it as easy as possible for wikis to migrate from Fandom. [21:22:48] [1/2] Yeah that's what led me to think the rate of false positives is too high. [21:22:48] [2/2] Could be that people are sharing PII but most are not important enough to act upon (e.g. I live in X city in Y country) [21:23:20] Or using offensive language which isn't bad enough T&S [21:26:14] Right person, right information, right time [21:26:35] At the right pace too if you're being overwhelmed [21:26:53] All the fun things you get to think about when making decision support software [21:27:33] well you can just have AI slop of course 🤔 [21:29:38] I am literally spending millions of pounds of public money to avoid that [21:30:20] Good [21:32:08] I don't even know what the full value of all of our work is [21:32:14] It's over a decade of research [21:33:25] not all were reviewed in the window [21:33:55] a lot of them are false positives/being over sensitive but tbh a lot are not wrong in some respect [21:34:43] I wanted to get it done and I couldn't be bothered to make it interesting [21:35:50] for every 10 bits of slop produced I will eventually deslop 1 thing [21:36:21] judging by the wikiaosis github you'll be there a while [21:36:48] Was that maths slop? [21:36:53] if I had more time I'd do it manually but I have so much else going on and I ended up being the single pillar propping everything up [21:36:58] Cause if the slop is going up 10:1, that's not good [21:37:26] no it was just shitty maths [21:40:56] a lot of it would be solved by writing my own test cases but that is something I really cba to do on top of everything [21:42:24] That pains me even more [21:42:33] Testing is more important than developing [21:49:35] I just don’t have the time right now but things need doing [21:51:09] tbh I am going to slide back into a bug fixing and gradual improvement role but it takes time I don’t have much of lol [22:18:13] I'm sure you've been saying that for as long as I've known you [22:18:53] also why does the website in your bio still lead to a nonexistent site