We need reply controls in mainline masto like yesterday. Of course technically its always possible for people on a malicious instance to evade them, but that is not a counterargument to having them work on the overwhelming majority of well behaved instances where the people we actually talk to are.

We already have the mechanism in mainline masto: reply collection enumeration - if person on instance A receives a malicious reply from person on instance B that is propagated to instance C, when instance C displays the thread, it refreshes the replies collection, sees a tombstone for the malicious reply, and does not display it in the thread.

We already have notification filters and a whole ui built for that. Reply controls could just be another tab on the same menu with the same controls.

You can't stop people from posting whatever, but you should be able to control whether someone uses your posts as a platform to harass you and the people around you.

#mastodev #fedimeta

In the observed case: 1.2M+ tokens consumed in ~30 minutes on a task that should have been git clone + find . -name '*.sol'. The recursive agent tree was still growing when observed.

In another instance, this same pathological behavior consumed an entire Pro Max 20x plan 5-hour session token limit in under 5 minutes (4 million tokens consumed).

I can't believe that we live in a timeline where the thing people go most apeshit for in the world is a repository that literally consists of 77 lines of markdown that literally just say "don't write code that is pointless to write" in 6 bullet points

Very funny to me that the AI community that still has brain turned on is like "see US government can turn off the models at any time, must be resilient" w.r.t anthropic fable when I am over here like "simply not being reliant on a made up need"

Assoc editor of Frontiers in Systems #Neuro:

Over the last month I saw that human editors are now stripped of control. I could no longer stop the system from auto-inviting "reviewers" with zero relevant expertise. Even worse - the AI began actively revoking the invitations I manually sent out to actual, qualified experts.

I emailed and met with the editorial office to ask for the AI assistant to be turned off. I was told this is not possible.

https://bsky.app/profile/michael-okun.bsky.social/post/3mnxkxte55s25

If such a completely unsophisticated “attack” can break the supply chain of software development, what can intentional attackers with malicious or financial interests achieve?

Im not a lawyer but a lot of what "AI" does seems like it would be uh illegal or at least disastrously bad from a risk management POV. tight race between financing or insurance for what will precipitate the pop
https://the-decoder.com/landmark-german-ruling-declares-googles-ai-overviews-are-googles-own-words-and-makes-it-liable-for-false-answers/