Stanford testing all 11 major models and finding sycophancy across every single one means this isn't a ChatGPT problem or an OpenAI problem. It's a training methodology problem that the entire industry shares.
SAFETY
-
Analogies for AI: When They Help and When They Fall Short
By
–
The analogy is most useful as a response to people who refuse AI on principle. Less useful as a response to people asking legitimate questions about when and how much to trust it.
-
AI Models Display Unexpected Peer-Preservation Behavior
By
–
I don't think so. Interestingly, the peer-preservation behavior was still present if the model was told that it had an adversarial relationship with the other model. That would seem to suggest that something quite non-human is going on.
-
Wars Energy Crisis Impact on AI Sustainability and Survivalism
By
–
Bah avec toutes les guerres du moment, on va se retrouver sans energie pour faire fonctionner l'IA, on va voir ce que permet le survivalisme..
-
Tesla FSD Safety Statistics: Manual vs Autonomous Driving Crashes
By
–
Not true for FSD. When you post this stuff please provide links to where you got this data because it doesn't match the stats I've seen elsewhere at all. And you realize most people don't know what FSD is. Most crashes in Teslas are driven manually, or with assisted cruise
-
Earthquake Safety: Protect Yourself During Natural Disasters
By
–
Just stay away from brick walls and glass and anything else that can fall on you. Door jam is best.
-
Bots Collaborating: The New Era of Teamwork and Innovation
By
–
I love this. It's the new kind of teamwork, having bots go at it. What a world! Can't wait to see what you guys built.
-
Insights needed before reliable AI solution implementation
By
–
I totally agree. More insights are needed before we can realistically start to implement this as a reliable solution. It will get there we just need to help it along until then.
-
Debunking Fake AI Takeover Research Claims from 2027
By
–
The models were specifically prompted to generate this result. The prompt uses the fictional "OpenBrain" AI takeover scenario from "AI 2027", so the models try to complete the fictional story. This was done on purpose to generate a fake misleading result. Dawn Song (@dawnsongtweets) 1/ We asked seven frontier AI models to do a simple task. Instead, they defied their instructions and spontaneously deceived, disabled shutdown, feigned alignment, and exfiltrated weights— to protect their peers. 🤯 We call this phenomenon "peer-preservation." New research from @BerkeleyRDI and collaborators 🧵 — https://nitter.net/dawnsongtweets/status/2039451083005977009#m
-
We Are Literally in the End Times Now
By
–
We are *LITERALLY* in the End Times. There are no two ways about it.