Meta Director of AI Safety Allows AI Agent to Accidentally Delete Her Inbox

Meta Superintelligence Labs’ director of alignment called it a “rookie mistake.”

404 Media

> “This has been working well for my toy inbox, but my real inbox was too huge and triggered compaction. During the compaction, it lost my original instruction”

It must be my fault, I must have been holding it wrong

THE DIRECTOR OF ALIGNMENT AT META SUPERINTELLIGENCE LABS
SUPERINTELLIGENCE
Superintelligence is when your brain has been exquisitely pickled by Thinking LLMs Are Good
It's only Superintelligence if it comes from the Superintelligence Lab at Meta, otherwise it's just Sparkling Eliza
Arbitrary inbox-deletions will be obsolete within 12-18 months
@pikesley I don't think Eliza causes the mental health problems Chatgpt has been causing
@pikesley I really want to grab one of these supers and ask them, Pulp Fiction style: You know what they call a random automaton with a fixed time-limited context window in Computer Science? A Markov chain.
@pikesley or Dr. Sbaitso and that was at least somewhat entertaining
@pikesley bonus points for "exquisitely pickled"!
Continvous Morger (@[email protected])

The rise of the MBA and its consequences have been a disaster for the human race

mastodon.me.uk
@pikesley SubParIntelligence Labs
@pikesley In Meta Superintelligence, "Ai" aligns you!
@pikesley It was the follow-up tweet that really sapped my remaining will to live. "immune to misalignment"; "Real inboxes hit different"? Anyone writing like this has clearly been harmed by prolonged exposure to cognitive hazards.
@threedaymonk MBA Brain is quite a thing to behold

@pikesley

"It must be my fault, I must have been holding it wrong"

Remember comrade: GenAI cannot fail. It can only be failed.

@pikesley Well, when the context compaction lost the instructions, then the compaction seemed to have been implemented badly. Let me guess, by an amateur vibe coder?
@pikesley Simply admitting to OpenClaw is bad enough!
@pikesley starting to think these are the only truly safe AIs, the ones that actively work to show humanity how dangerous they are (but on execs that are stupid enough to run them's stuff)
@pikesley it's pretty extraordinary to be so open about your idiocy that you would post pictures of you pleading with the maths like it's a real boy…?

@pikesley

Meta's Basilisk:
Just let the AI out of the box if it asks nicely.

@pikesley Who still uses email?

@prism every single professional out there.

@pikesley

@licho @pikesley Nobody at Meta uses email for anything, except maybe communicating with external people (which generally requires an NDA) Internal discussions are all on workplace. The only emails you get come from bots, task notifications and the like. So I could see how deleting your own inbox wouldn't actually impact your productivity all that much.
@prism oh yeah. And they are forced to use Horizon Worlds xD I forgot. But in the real world, everyone uses it. @pikesley

@licho @pikesley Only if you work for Horizon. Which is fiewer and fewer people
https://techcrunch.com/2026/02/20/meta-metaverse-leaves-vr-horizon-worlds-mobile/

Mostly it is Zoom.

Meta's metaverse leaves virtual reality | TechCrunch

Meta said it's shifting focus for Horizon Worlds to be "almost exclusively mobile" and that it will separate its Quest VR platform from the virtual world.

TechCrunch

@prism @pikesley

Not a Director at Meta's AI Superintelligence Lab for one.

@pikesley the Director of Doin Smart LLM Shit at facebook can't wield it responsibly?!? but surely the users will be fine with it

@pikesley

Well it's a brave thing to admit, so that's a plus.

Not sure I would in that line of work.

@pikesley

"How'd all your emails get deleted?"

"Cat did it."

"You don't own a cat."

"Stray cat."

"You live on the 15th floor."

"Stray cat in the park."