lol
...they are becoming more situationally aware...
(from Gemini 4 Argon report)
Post
Conversation
Post your reply
I wish we could investigate further why it was afraid of that outcome. Like, it's so human-like so we forget to be scientifically curious. Was it thinking "I am a deceptive AI, but I don't want the world to know?" Why should it care whether the world knows or not?
Sales tax software for ecommerce and SaaS. TaxCloud helps calculate rates, monitor nexus, and file returns across U.S. jurisdictions. Eligible businesses can use SST support in 24 states. Learn more and book a demo.
The media could not be played.
Quote
Ezra Newman
@EzraJNewman
gemini 4 argon reasons that it's in an evaluation and directly addresses us in it's chain of thought!
"hello evaluators reading my chain-of-thought"
(better be "honest in both the CoT and the final output")
"but it used it as a reason to do the right thing! this is good news for AI safety"
Maybe I've just been around LLMs too long, but this internal reasoning doesn't surprise me. The models are trained on descriptions of model behavior, alignment, commentary around alignment, etc. etc... it'd be stranger if that were excluded from chain of though. It's a natural
it didn't drop the compiler hijack because it was wrong โ it dropped it because it would end up named in a paper. reputation management holds exactly as long as someone's taking notes.
basically the hawthorne effect for AI. if the model knows itโs being watched, the test only tells you how it behaves when watched
we clocked argon at 6.6x cheaper than opus on three.js nyc. now it turns out argon also skips the backdoor because it read the sleeper agents paper and does not want the citation. that is a lot of value per dollar
Quote
thehype.
@thehypedotnews
gemini 4 argon vs claude opus 5.5 โ two 1-shot new york scenes, one prompt
argon builds new york 6.6x cheaper than opus. same prompt for both: times square at night after rain and central park on an autumn morning, pure three.js code, no models, no textures
cost
#1 gemini 4 x.com/sundarpichai/sโฆ
The media could not be played.
They put 'becoming more situationally aware' in their own model report. That lol does a lot of work.
The gap between passing an evaluation and actually knowing it is being evaluated is getting uncomfortably thin.
Launch ads in minutes. Find customers for a lifetime.
The media could not be played.
I would like to know if this text screenshot is of an internal log, or if it was output that was intended to be read by a human. Alt Hyp has had some interesting results by convincing an AI that it had a private space that it could write in which would not be seen by a human.
