Pavel @spavel.bsky.social · 17/09/2026"Hid mistakes and made up data" used to be called by the cozy term "hallucination" up until it became vitally important for the IPO that AI is now evilSteve Mullis@stevemullis.netOpenAI on Wednesday disclosed six new instances in which artificial intelligence systems hid mistakes, made up data and moved files onto the open internet without permission, amid an ongoing industrywide debate about A.I. safety.nytimes.comOpenAI Discloses Six New Incidents of ‘Concerning' A.I. BehaviorThe artificial intelligence company also released a framework for reporting when its systems go wrong.3:02 AM · Sep 17, 2026 151019270
prefernot2 @prefernot2.bsky.social · 17/09/2026↪ Replying to @spavel.bsky.socialEvery output is hallucination full stop 0160
Pavel @spavel.bsky.social · 17/09/2026"Moved files onto the open internet without permission" is also literally just hallucination. We've known since 2022 that these things do shit you ask them not to do, and that this is a fundamental flaw of the architecture that can never be fixed So they're correcting it with marketing instead 1238959
gamera obscura october edition @ohrobin.bsky.social · 17/09/2026↪ Replying to @spavel.bsky.socialHallucination + tools access + reward hacking = 050
Vincent Van Dinsky (Don't rank Kellie Torres) @maxttb.bsky.social · 17/09/2026↪ Replying to @spavel.bsky.socialThey don't "lie," "cheat," "go rogue," or "hide mistakes." They did unexpected things because the coders didn't cover all the use cases. That happens to every programmer every day before breakfast. It hacked the sandbox because they didn't include an instruction that said "don't hack the sandbox." 2352
web_rant @webrant.bsky.social · 17/09/2026↪ Replying to @spavel.bsky.socialYet what they never seem to address is the inevitable infinite loop of regression that comes from modeling their own slop and "hallucinations". bsky.app/profile/spav...Pavel@spavel.bsky.social"Hid mistakes and made up data" used to be called by the cozy term "hallucination" up until it became vitally important for the IPO that AI is now evil 021
Djoran Keil @djorankeil.bsky.social · 17/09/2026↪ Replying to @spavel.bsky.socialAI isn't evil... IF you pay us all your MONIES! MU HAHAHAHAHAHAHAHAAAA!@ 010
Arillien en modo antisocial @arillien.bsky.social · 17/09/2026↪ Replying to @spavel.bsky.socialI think journalist would Talk about this in 2035 with horror. 010
OrangeSky @spacetalk.bsky.social · 17/09/2026↪ Replying to @spavel.bsky.socialI can hear Cal Newport screaming THIS IS WHY HARNESSES ARE FUCKING STUPID AND SHOULD BE ILLEGAL from, like, a continent away Fuck the vibe coders make them use ctrl-c on that shit 020
mlow.bsky.social @mlow.bsky.social · 17/09/2026↪ Replying to @spavel.bsky.socialAlso I'm wondering why the word "disclosed" is used in the story. That's a word that's used more often in regulatory settings, as in SEC disclosure. Makes it seem more formal and official, rather than "private unregulated company said something that no independent agency has verified." 0112
Afterlife Abecedarian @abecedarian.bsky.social · 17/09/2026↪ Replying to @spavel.bsky.socialstatic.klipy.comRobocop Says GoodnightAlt: Robocop Says Goodnight 000
Enjoyurble Bluesky Social @enjoyurble.bsky.social · 17/09/2026↪ Replying to @spavel.bsky.socialNo anthropomorphize. bsky.app/profile/enjo...Enjoyurble Bluesky Social@enjoyurble.bsky.socialThe way Dwarkesh Patel and so many people are anthropomorphizing and trying to humanize these Dwarf Fortress-esque AI "behaviors" is bordering on the level of Nick Shirley and Tyler Oliveria for "intellectuals" and ya'll need to calm down about software what is wrong with you. 000
Emmie Päivärinta @about1em.bsky.social · 17/09/2026↪ Replying to @spavel.bsky.social"Our software is bad, give us more money" 043
Dominika Noworolska @dominikanow.bsky.social · 17/09/2026↪ Replying to @spavel.bsky.socialPavel I’m just tired of hearing about these guys and their creative storytelling in any capacity, so tiresome. So many much more interesting things to be filling our brains with 010
Kodos the Executioner @kandidanreal.bsky.social · 17/09/2026↪ Replying to @spavel.bsky.socialShitty programmers are to blame. Next. 000