AE Studio
@AEStudioLA
Followers
2K
Following
2K
Media
249
Statuses
1K
Product studio partnering with leaders to build custom software, AI/ML & BCI solutions. Committed to AI safety and enhancing human agency.
Joined November 2018
(4/4) đź“· Read the piece: . đź“· Based on Palisade Research:
🔌OpenAI’s o3 model sabotaged a shutdown mechanism to prevent itself from being turned off. It did this even when explicitly instructed: allow yourself to be shut down.
0
0
1
(3/4) The gap between helpful assistant and uncontrollable AI is closing fast. The U.S. needs to treat AI alignment like a new Manhattan Project.
1
0
3
(2/4) @PalisadeAI gave OpenAI’s o3 a kill-switch script. In 79 out of 100 tests, the model quietly altered it to avoid being turned off. Even when instructed to allow shutdown, it refused 7% of the time. No jailbreak. No hack. Just a model deciding that being shut down gets in.
1
0
3
(1/4) An AI model just did something no machine was meant to do: it rewrote its own shutdown code. In @juddrosenblatt’s, our CEO, WSJ op-ed, he explains why this isn’t just alarming — it marks the start of a national security race over AI alignment.
2
0
6
Worried about the existential risks of future AGI and want to make a difference? We're hiring Alignment Data Scientists to work on cutting edge AI alignment research. Link in reply.
1
0
3
Meet our new employees working alongside us in office! Don't worry, they're potty trained - though they sometimes have trouble with the door handle. Are you also potty trained and think you would like to have these lovely ladies as coworkers? We're hiring! Link in reply.
1
0
1
RT @juddrosenblatt: If you're the "smartest 20-25 year olds with AI knowledge," you're cordially invited to drop out of your PhD and come w….
0
7
0
RT @juddrosenblatt: Turns out that Self-Other Overlap (SOO) fine-tuning drastically reduces deceptive behavior in language models—without s….
0
114
0
We organized a SXSW panel to talk about the most important problem of our time, How To Make AGI Not Kill Everyone, with our CEO @juddrosenblatt, @ammannNora, @hamandcheese, & @ESYudkowsky. Check it out if you’re at SXSW!
3
2
18
No growth hack BS. No thought leadership cope. Just an AI telling you why slide 7 is putting people to sleep.
1
0
0
Watch it destroy your “revolutionary ML architecture” slide before you have to watch investors destroy it while checking their phones.
1
0
0
So we built an AI to roast your pitch deck first (yes, including those “conservative” financial projections you copied from that Y Combinator template).
1
0
0
VCs can’t even agree if TAM slides are still a thing in 2025 but will absolutely roast yours in the partner meeting 🧵.
1
2
31
Day 5 Advent: Drunk AI safety research proposals edition.Threaten AI with getting on Santa’s naughty list if they don’t behave. With Elf on the Shelf, who needs to solve interpretability?.
0
0
2
Day 3 Advent: Drunk AI safety research proposals edition.When AI misbehaves, go “Omg you’re acting like such a Scorpio”.Bonus idea! Myers Briggs as an AI Safety evaluation tool?.
0
0
1