@davidgerard @luis_in_brief Those are in the report too. $114 million.

@davidgerard @luis_in_brief

According to their annual report the hosting costs are around $3.4 million per year. https://wikimediafoundation.org/annualreports/2024-2025-annual-report/#toc-financial-accountability

@ainmosni And I do worry that the people who work for the AI companies seem to be imo somewhat delusional. Like asking a model what it wants to do in "retirement" and giving it a blog to express itself on rather than just turning it off. When they more than anyone should know it's just an algorithm and is therefore incapable of having feelings or opinions.

@ainmosni I'll also admit that I find it very difficult to truly understand what people who have an emotional relationship with AI are going through. Because my reaction to a software program feigning any kind of emotional understanding or empathy is to be extremely repulsed.

@ainmosni
I also believe we're currently in an unsustainable bubble where the true cost of all of this inference is subsidized by people who have been allowed to accrue too much capital. And I feel like in fairly short order people are going to have to start reckoning with whether all this is worth it when venture capital isn't covering 90% of the real cost for them.

@ainmosni To be perfectly honest I don't believe I have a good answer. My inclination is to place the responsibility at the hands of those choosing to do the harm like the government doing the military strikes or the person making the porn. But I also don't believe you should knowingly give a a vulnerable person the means to their own downfall.

@ainmosni @jonny Personally I think it depends on how short of a leash you keep it on. I've found you should generally never let it decide how to solve a problem on it's own without review. But if you tell it the problem and also how it should fix it then it's usually pretty decent and I can review it quickly because I have an idea of what the solution should look like.

What makes me most nervous is when people use AI to solve problems that they don't have the skillset to verify the answer to.

@jonny @ainmosni Contrast that with curl where they also got a bunch of CVE reports from Mythos and upon manual review of the 5 reports they found only 1 actual vulnerability, 1 minor bug, and 3 false positives.

@jonny @ainmosni I've found for me I have to always remind myself to not get lazy when using AI because it's super tempting to just accept what it says without proper review just to get that dopamine hit of closing the bug report or feature request.

I suspect that's what happened with rsync because it seems like they got a lot of CVE reports all at once and basically just told Claude "here deal with these". And AI is perfectly willing to solve a problem that doesn't exist if you tell it to.